acceptodds
Under review as a conference paper at ICLR 2027

Beyond Flat Spaces: Hyperbolic Manifolds for Unexploded Ordnance Identification

Abstract

Unexploded Ordnance (UXO) identification is a safety-critical visual recognition problem characterized by extreme data scarcity and substantial visual similarity across ordnance categories. We analyse UXO embeddings using complementary geometric diagnostics and find evidence of low effective dimensionality and stronger tree-like structure relative to a calibrated null model. Motivated by this evidence, we introduce LorentzTIM, a transductive information-maximization method formulated directly on the Lorentz hyperboloid. LorentzTIM represents samples and class prototypes in hyperbolic space and performs transductive prototype adaptation using geodesic distances and Riemannian optimization. On the CTX-UXO benchmark, LorentzTIM consistently improves F1 over the evaluated Euclidean and hyperbolic baselines, with consistent gains across all shots regime. The improvements persist across different feature extractors and extend to CUB-200-2011, FGVC-Aircraft, and tieredImageNet, supporting the broader applicability of the proposed approach beyond UXO identification.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI for social good, few-shot learning, hyperbolic geometry, transductive learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.80 over 5 reviews, above 27% of the others.
Reviewer_VMSGOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2(+1)
Read review
Paper summary

This submission deals with the problem of identifying UXO: unexploded ordnance (ammunition that did not explode when originally fired, and may explode unexpectedly at some time in the present). The authors claim that UXO identification using Euclidean features is especially hard because of data scarcity and high intra-class variance even despite strong inter-class similarity. To address this, the paper proposes using hyperbolic features instead. It first studies the structure and dimensionality of UXO representations to argue that they exhibit a geometry resembling a hyperbolic space, compared to a null baseline. Using this motivation, it then proposes a transductive information maximization (TIM) strategy built on top of embeddings lifted into a hyperbolic space. The method, named LorentzTIM, implemented over frozen DINOv3 features, is evaluated on a benchmark for comparison against similar baselines.

Critical strengths

The experiments in the paper appear correctly implemented; the authors add several baselines to the rebuttal, which helps compare different classifiers on top of frozen image embeddings.

Critical weaknesses

W1. The geometric motivation is not empirically established.

The proposed method, LorentzTIM, introduces significant complexity (§4.2) motivated by the claim that UXO embeddings are "tree-like". But why are UXO embeddings special in this regard, i.e., in the presence of large foundation vision models, how can one motivate the development of considerable algorithmic machinery to tackle the specific problem of UXO identification, and how generalizable are the paper’s claims beyond the task of UXO identification? In addition,

  • Table 2 (row 1 in particular) seeks to establish that UXO embeddings are “tree-like”. I disagree that the current result supports this claim: the Gaussian control lacks class structure and is too weak of a control, therefore exhibiting higher tree-likeness in comparison to this baseline is not sufficiently informative. At minimum, a comparison with real image embeddings (e.g. Imagenet-1k) would be relevant.

  • Likewise, no evidence in the paper supports lines 425-426: “when the data are more intrinsically tree-like, hyperbolic methods tend to outperform flat Euclidean baselines.” This would require a correlation between the “tree-likeness” of embeddings and the advantage of LorentzTIM over the baselines; given that the paper only uses one dataset and one model in the current manuscript, it is unclear to me how this conclusion can be made from the current evidence. Even with the three additional datasets promised in the penultimate section (which are, importantly, missing), we would only have four data points in total. This would be far too small a sample size for making a general claim about the correlation between the geometry of the embeddings and the effectiveness of LorentzTIM on said embeddings.

W2. Very limited empirical strength towards the claimed advantages of LorentzTIM.

  • The comparison between TIM over Euclidean and hyperbolic embeddings is not sufficiently controlled. For instance, LorentzTIM uses Eqn. 21 as a regularizer. It is unclear to me whether this regularizer was used for the Euclidean version as well, but from the manuscript, the answer seems no. Likewise, the ablation in Table 4(a) does not satisfyingly isolate the benefit of hyperbolic embeddings: it looks like the curvature was set to zero while keeping the other hyperparameters at their LorentzTIM settings, which is potentially suboptimal for Euclidean TIM. As such, I find it hard to understand whether the 2-4 pp. gains from LorentzTIM (vs. tuned Euclidean TIM) come from the hyperbolic treatment or an improved TIM itself.  

  • In particular, if the core claim is that UXO embeddings benefit from being lifted into a hyperbolic space, the benefit should plausibly occur independently of applying TIM. A simple experiment would compare linear and nonlinear classifiers trained over Euclidean and hyperbolic embeddings, with matched hyperparameter tuning budgets. This is perhaps attempted in SimpleShot vs Hyp-SimpleShot: Hyp-SimpleShot there has rather small improvements compared to SimpleShot (all less than 1 pp). Perhaps hyperbolic embeddings benefit especially from TIM, but the paper neither (i) provides an ablation showing the general superiority of hyperbolic embeddings nor (ii) clarifies the qualification that hyperbolic embeddings work well only under the specific, proposed modifications. If (ii) reduces to a really narrow setting (hyperparameter configuration, dataset and model), that brings us back to the generalizability issue from W1.

  • The only feature extractor presented in the manuscript is DINOv3; there is no justification as to why this model is suitable for the study, and no other models are considered. Even though the paper mentions CLIP was also considered in the Appendix, said Appendix is missing. Similarly, the three datasets promised in lines 471-473 are missing (evidently, because they were in the missing Appendix). Figure 5 is missing the baseline numbers, therefore, it currently remains inconclusive towards a proper assessment of the advantage of LorentzTIM. More broadly, though, even with DINOv3 and a single CLIP model, the study would have been far too limited for making a general claim about vision features.

Questions for authors

Please treat the points mentioned under Weaknesses as critical questions. In addition, for a potential re-assessment of my score,

Q1. I would appreciate seeing more ablations: in particular, a more direct ablation comparing Euclidean and hyperbolic embeddings under various downstream models (linear classifier, 2-layer MLP, and finetuning) would be really helpful.

Q2. Can the authors motivate why the task of UXO is especially suited to the proposed LorentzTIM? What do the current results tell us about which tasks will benefit from the proposed method? To make this question more concrete for the rebuttal, I would appreciate seeing results on the other three datasets from the missing appendix.

6 comments
AuthorsOpenReview ↗

Response to Reviewer VMSG (1/4)

We thank the reviewer for the detailed feedback. The CLIP results and the three further datasets were in Appendix G (Tables 9-12), which was submitted as supplementary material. We report them directly below (Tables A and D).

Summary. (1) Controls that isolate the geometry within each method: Euclidean TIM with the regulariser and its own search, and linear and 2-layer MLP classifiers in both geometries.

(2) Results with CLIP and on three further datasets, against both a TIM with the same hyperparameters and a separately tuned TIM.

AuthorsOpenReview ↗

Response to Reviewer VMSG (2/4)

W1. Geometric motivation.

  • Why geometry, given foundation models. LorentzTIM does not replace foundation models. It adapts class prototypes on their frozen features, and it improves on Euclidean TIM with both backbones we tested (Table A). The added cost is small: 197-199 ms per 5-way 5-shot episode on one RTX 4090 (Table 8);

  • Is UXO special? UXO motivated the work through scarce labels, a long tail, environmental degradation and visual similarity across categories, but the method is not UXO-specific. The same gains appear on CUB-200-2011, FGVC-Aircraft and tieredImageNet (Table D under W2);

  • Strength of the control and comparison with real image embeddings. We agree that the original Gaussian control has no class structure. We added a class-conditional and a full-covariance control, both generated in the full feature space and L2-normalised like Z (Table B). The UXO embeddings Z have the lowest δ_rel (.2453) and the highest cophenetic correlation (.768 / .613) of the four point sets. The class-conditional control comes close (δ_rel .2678, with overlapping confidence intervals; cophenetic .683 / .567), whereas the two controls without class structure are clearly less hierarchical (δ_rel .3070 and .4215; cophenetic at most .520 / .454). The tree-likeness of Z is therefore explained largely by its class structure. Following the suggestion, we also computed the diagnostics on ImageNet-derived embeddings (tieredImageNet): δ_rel is .147 with DINOv3 and .198 with CLIP, against .2453 and .284 for CTX-UXO. Natural-image embeddings are therefore also tree-like, which is consistent with LorentzTIM improving on all three additional datasets. We present tree-likeness as a general motivation for hyperbolic adaptation rather than as a property specific to UXO.

Table A. CTX-UXO, macro-F1 (%), 1,000 paired test episodes.

BackboneMethod1-shot3-shot5-shot10-shot
DINOv3TIM (tuned)49.5465.1869.3576.14
LorentzTIM (tuned)53.8367.1072.1080.51
CLIPTIM (same ψ as LorentzTIM)34.4643.0747.5252.52
TIM (own random search)34.8243.4449.8555.71
LorentzTIM (tuned)39.2946.7851.4256.57

Table B. Geometric diagnostics for the UXO embeddings and three controls (DINOv3).

DataParticipation ratioδ_rel (tree-likeness), 95% CIρ_coph (Pearson / Spearman)
UXO embeddings Z29.04.2453 [.213, .274].768 / .613
Single-Gaussian control (Eq. 11)12.88.4215 [.394, .462].520 / .454
Full-covariance Gaussian, no class structure (new)35.29.3070 [.255, .355].505 / .411
Class-conditional Gaussian (new)34.44.2678 [.233, .318].683 / .567
AuthorsOpenReview ↗

Response to Reviewer VMSG (3/4)

W2. Controls and empirical strength.

  • Was the regulariser of Eq. 21 used for Euclidean TIM? We added Euclidean TIM with the regulariser, tuned by its own random search with the same budget (Table C). Adding the regulariser changes Euclidean TIM by −0.82 / +0.67 / +1.38 / −0.81 points; LorentzTIM remains above it at every shot (by 5.11 / 1.25 / 1.37 / 5.18 points).

  • Table 4(a) keeps the LorentzTIM hyperparameters. This is correct, and the tuned Euclidean variant with the regulariser above is the comparison the reviewer asks for. In addition, the fixed-parameter row of Table 3 uses TIM's own hyperparameters, not ours, and still exceeds tuned TIM at every shot (+1.74 / +1.26 / +2.32 / +3.43).

  • Hyperbolic lifting alone. In the balanced protocol, Hyp-SimpleShot exceeds SimpleShot at every shot, by +0.69 / +0.57 / +0.69 / +0.41 points, while LorentzTIM exceeds tuned TIM by +4.29 / +1.92 / +2.75 / +4.37. Our claim is therefore the narrower one, option (ii) in the review: hyperbolic geometry helps in combination with transductive prototype adaptation.

  • Choice of DINOv3. Its self-supervised features are strong on fine-grained recognition with frozen backbones (Siméoni et al., 2025), which matches the UXO setting. CLIP is evaluated as a second backbone (Table A). The paper reports results for these two backbones and does not extend the claim to other vision features.

  • Three further datasets. LorentzTIM improves on Euclidean TIM run with the same hyperparameters in all 24 dataset × backbone × shot settings, by +1.31 to +3.67 points, and on the separately tuned TIM in all 24, by +1.34 to +3.43 points (Table D).

  • Figure 5. The baseline values behind Figure 5 are given in Table D.

Table C. CTX-UXO, DINOv3, macro-F1 (%), 1,000 paired test episodes.

Method1-shot3-shot5-shot10-shot
TIM, tuned by random search49.5465.1869.3576.14
TIM + origin regulariser (Eq. 21), tuned by random search48.7265.8570.7375.33
LorentzTIM, c → 0 (Table 4(a); with regulariser, LorentzTIM settings)48.8664.7168.8275.58
LorentzTIM, TIM's fixed hyperparameters (no search)51.2866.4471.6779.57
LorentzTIM, tuned53.8367.1072.1080.51

Table D. Values behind Figure 5, with both TIM protocols: macro-F1 (%) with DINOv3, 1,000 paired episodes per setting.

DatasetMethod1-shot3-shot5-shot10-shot
CUB-200-2011TIM (same ψ as LorentzTIM)83.6689.8692.1493.95
TIM (own random search)83.8390.3691.9093.88
LorentzTIM86.2491.7093.8295.26
FGVC-AircraftTIM (same ψ as LorentzTIM)80.4687.9390.7393.11
TIM (own random search)80.6688.5390.4493.03
LorentzTIM83.2090.1292.6594.78
tieredImageNetTIM (same ψ as LorentzTIM)78.8686.1288.9991.52
TIM (own random search)79.0886.8188.6591.42
LorentzTIM82.4188.7791.2293.36

Q1. Downstream classifiers.

We added the ablation in both geometries for two downstream models: a linear classifier (Euclidean logistic regression vs hyperbolic multinomial logistic regression) and a 2-layer MLP (ReLU hidden layer of 32-128 units; the hyperbolic version applies the hidden layer in the tangent space at the origin and uses the hyperbolic MLR as output layer). Each is trained on the support set with its own random search and the same budget (Table E). Together with SimpleShot versus Hyp-SimpleShot (inductive) and T = 0 (Table 4(b)), this separates the geometry from the transductive step.

Table E. CTX-UXO, DINOv3, macro-F1 (%).

Method1-shot3-shot5-shot10-shot
Euclidean logistic regression (support set, tuned)46.7061.0467.7975.20
Hyperbolic MLR (support set, tuned)47.5761.4267.1975.07
Euclidean 2-layer MLP (support set, tuned)46.7560.9967.1674.14
Hyperbolic 2-layer MLP (support set, tuned)46.6561.5367.5674.31
SimpleShot46.9060.5066.4572.99
Hyp-SimpleShot47.5961.0767.1473.40
LorentzTIM53.8367.1072.1080.51

With matched tuning budgets, the hyperbolic linear classifier shows no consistent advantage over the Euclidean one (+0.87 / +0.38 / −0.60 / −0.13 points at 1 / 3 / 5 / 10 shots), and neither does the hyperbolic MLP (−0.10 / +0.54 / +0.40 / +0.17 points). This matches option (ii) in the review: the hyperbolic advantage comes with transductive adaptation and is not a general property of the embeddings. We did not fine-tune the backbone: LorentzTIM is a training-free method over frozen features, and fine-tuning would change the embeddings whose geometry is being compared. All classifiers in Table E and all transductive methods operate on the same frozen features.

Reviewer_gF3FOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LorentzTIM for few-shot identification of unexploded ordnance. Frozen image features are reduced with PCA and mapped to the Lorentz model. Class prototypes are initialised from the support images and updated using a transductive information-maximisation objective. The authors report higher macro F1 than Euclidean TIM on CTX-UXO, with additional results using CLIP and on three other datasets.

Critical strengths

The application is well motivated. Labelled UXO images are scarce, and different ordnance types can look similar. The method works with a frozen image encoder.

The gains over TIM are consistent across the four shot settings, both with fixed parameters and when both methods are tuned.

The methods are evaluated on 1,000 shared episodes, with confidence intervals and paired tests against original TIM.

The paper provides ablation studies. Table 4 examines the effects of curvature, prototype updates, loss terms, normalisation and the optimiser. The drops from removing curvature and prototype updates show that these components contribute to the reported performance.

Critical weaknesses

Check the question, please.

Questions for authors

My main concern is the novelty. LorentzTIM moves the established TIM objective to the Lorentz model using standard centroids and Riemannian Adam. This gives consistent gains, but I am not convinced the current paper provides enough new insights as well. Gao et al. (2021) already study hyperbolic few-shot learning in a transductive setting, using a different, meta-trained approach. The paper needs more study on why or when the proposed method works.

I do not see how the control in Eq. 11 rules out class clustering, as claimed at L270–272. It samples from a single Gaussian based on the leading covariance directions and variances, but is not constructed to preserve the separate class clusters. The difference in Table 2 could therefore reflect clustering rather than a multilevel hierarchy. This leaves the geometric motivation less convincing than the text suggests.

Is Figure 4b comparing the method with ordinary Tim? are there any imbalance experiments with α-TIM?

L428–431 attributes the gains mainly to visual similarity rather than class frequency. Figure 4c only reports per-class F1; it neither shows class frequencies nor measures similarity. The plot does not establish this explanation. Please provide the analysis behind the statement or describe it as a possible explanation.

Minor comments

Please use a clearly labelled 0–100 scale for F1 scores, for example 53.83 ± 0.94 instead of .5383 ± .0094.

Table 2 needs a short explanation of how to interpret each diagnostic, especially the control entry “29.04/12.88”.

L190–196 introduces optimiser details before the prototype initialisation and objective. I would find the method easier to follow if these details came after the objective.

At L324, Eq. 19 defines the entropies. The objective is Eq. 17, or Eq. 21 with regularisation.

At L69–70, move the citations to the end of the sentence. Figure 1 would fit better in the method section; Figure 2 could introduce the problem earlier.

The related work would benefit from discussing Moreira et al. (WACV 2024) on hyperbolic versus Euclidean few-shot embeddings. Also, there is a lot of missing literature. Citing a survey would help: Hyperbolic Deep Learning in Computer Vision: A Survey by Mettes, Ghadimi Atigh, Keller-Ressel, Gu and Yeung.

4 comments
AuthorsOpenReview ↗

Response to Reviewer gF3F (1/4)

We are grateful to the reviewer for the constructive evaluation. We respond to each question below.

Summary.

(1) The contribution and the evidence on why and when LorentzTIM helps are stated explicitly.

(2) Two stronger geometric controls, which show that the tree-likeness of the UXO embeddings comes largely from their class structure. (

(3) An imbalance comparison with tuned TIM and α-TIM.

AuthorsOpenReview ↗

Response to Reviewer gF3F (2/4)

Q1. Novelty, and why and when LorentzTIM works.

We agree that transductive hyperbolic few-shot learning predates our work. Gao et al. (2021) meta-train an encoder that generates a task-aware curvature; our contribution is a training-free transductive formulation on the Lorentz model over frozen features, in which the class prototypes are adapted on the hyperboloid. On why and when it works, the results support three statements (Table A):

  • Hyperbolic lifting alone gives a small but consistent gain in the balanced protocol. Hyp-SimpleShot exceeds SimpleShot at every shot, by +0.41 to +0.69 points.

  • The gain appears when prototypes are adapted on the manifold using the query set. LorentzTIM exceeds tuned Euclidean TIM by +1.92 to +4.37 points. Removing adaptation (T = 0) lowers 5-shot macro-F1 from 72.10 to 67.14, which is exactly Hyp-SimpleShot.

  • The gain is not an artefact of tuning or of the regulariser. With TIM's fixed hyperparameters (ω = 0), LorentzTIM still exceeds tuned TIM at every shot, whereas the c → 0 variant of Table 4(a) (K = −10⁻⁵, ψ*_S unchanged), which keeps the regulariser, falls below tuned TIM at every shot.

Mechanism. On the hyperboloid (K = −1), the geodesic distance between two points at radii r₁, r₂ separated by an angle θ satisfies cosh d = cosh r₁ cosh r₂ − sinh r₁ sinh r₂ cos θ. Near the origin it agrees with the Euclidean distance to first order; away from it, the two differ by a curvature term that depends non-linearly on both radii and on θ. For two points at the same radius r and a fixed angle, the Euclidean distance grows linearly with r, whereas the hyperbolic distance grows faster: for θ = 0.5, their ratio is 1.04 at r = 0.5, 1.13 at r = 0.9 and 1.27 at r = 1.3. Through the posterior (Eq. 15) and the mutual-information objective (Eq. 19), the Riemannian updates therefore act on a distance whose dependence on the radial position of the prototypes differs from the Euclidean one, which changes the direction and the size of the prototype updates. The c → 0 control of Table 4(a) removes only this curvature: with the regulariser, the search and the initialisation unchanged, it costs 2.39 to 4.97 points and drops below tuned TIM at every shot. The lifted inputs lie at geodesic radius 0.85 to 0.89, and during adaptation the prototypes move outwards at 3, 5 and 10 shots, to median radii 2.24, 1.26 and 0.64 (reported to Reviewer DCaD, Q2); at 1 shot they move inwards, from 0.89 to 0.31, so the curvature acts along the trajectory. At these radii the curvature term is substantial: for two prototypes at radius 2.24 (1.26) separated by 0.5 rad, the geodesic distance is 1.77 (1.25) times the Euclidean one. The gain therefore depends on the curvature during adaptation, as the c → 0 control shows. Table 4(c, d) complements this by quantifying the contribution of the marginal-entropy and mutual-information terms.

Table A. Differences in macro-F1 (percentage points), CTX-UXO, DINOv3.

Comparison (percentage points)1-shot3-shot5-shot10-shot
Hyp-SimpleShot − SimpleShot (lifting only)+0.69+0.57+0.69+0.41
LorentzTIM − tuned TIM (adaptation on the hyperboloid)+4.29+1.92+2.75+4.37
LorentzTIM with TIM's fixed hyperparameters − tuned TIM+1.74+1.26+2.32+3.43
LorentzTIM with c → 0 (Table 4(a); same regulariser and settings) − tuned TIM−0.68−0.47−0.53−0.56

The same pattern holds beyond CTX-UXO, on three further datasets and with a second backbone (Table B).

Table B. LorentzTIM − Euclidean TIM, macro-F1 (percentage points), 1,000 paired episodes per setting. "Same ψ": TIM run with the hyperparameters selected for LorentzTIM (submitted Tables 9-12); "own search": TIM tuned by its own random search with the same budget. Holm-adjusted Wilcoxon p < 0.001 for every entry.

SettingLorentzTIM − Euclidean TIM1-shot3-shot5-shot10-shot
CTX-UXO, CLIPvs TIM, same ψ+4.83+3.71+3.90+4.05
vs TIM, own search+4.47+3.34+1.57+0.86
CUB-200-2011, DINOv3vs TIM, same ψ+2.58+1.84+1.68+1.31
vs TIM, own search+2.41+1.34+1.92+1.38
FGVC-Aircraft, DINOv3vs TIM, same ψ+2.74+2.19+1.92+1.67
vs TIM, own search+2.54+1.59+2.21+1.75
tieredImageNet, DINOv3vs TIM, same ψ+3.55+2.65+2.23+1.84
vs TIM, own search+3.33+1.96+2.57+1.94
CUB-200-2011, CLIPvs TIM, same ψ+3.14+2.15+1.76+1.34
vs TIM, own search+2.94+1.53+2.07+1.43
FGVC-Aircraft, CLIPvs TIM, same ψ+3.20+2.17+1.86+1.54
vs TIM, own search+2.95+1.41+2.23+1.65
tieredImageNet, CLIPvs TIM, same ψ+3.67+2.35+2.15+1.73
vs TIM, own search+3.43+1.60+2.53+1.84
AuthorsOpenReview ↗

Response to Reviewer gF3F (3/4)

Q2. Gaussian control and class clustering.

We added a full-covariance Gaussian and a class-conditional control, in which each class is sampled from a Gaussian with that class's mean and covariance, in the class proportions of Z. Both controls are generated in the full feature space and L2-normalised like Z, without truncation (Table C). The UXO embeddings Z have the lowest δ_rel (.2453) and the highest cophenetic correlation (.768 / .613) of the four point sets. The class-conditional control comes close (δ_rel .2678, with overlapping confidence intervals; cophenetic .683 / .567), whereas the two controls without class structure are clearly less hierarchical (δ_rel .3070 and .4215; cophenetic at most .520 / .454). The tree-likeness of Z is therefore explained largely by its class structure. We therefore withdraw the stronger claim at L270-272.

Table C. Geometric diagnostics for the UXO embeddings and three controls (DINOv3).

DataParticipation ratioδ_rel (tree-likeness), 95% CIρ_coph (Pearson / Spearman)
UXO embeddings Z29.04.2453 [.213, .274].768 / .613
Single-Gaussian control (Eq. 11)12.88.4215 [.394, .462].520 / .454
Full-covariance Gaussian, no class structure (new)35.29.3070 [.255, .355].505 / .411
Class-conditional Gaussian (new)34.44.2678 [.233, .318].683 / .567

Q3. TIM versus α-TIM under imbalance.

Fig. 4b ran LorentzTIM and Euclidean TIM with a shared default configuration (T = 50, learning rate 0.05, τ = 10, weight 0.1 on the mutual-information term, and ω = 0.01 for the regulariser of LorentzTIM, which Euclidean TIM does not have), with no search for either method. This default configuration is not the TIM configuration of the fixed-parameter row of Table 3. Fig. 4b also did not include α-TIM. Under this shared default configuration, Euclidean TIM reaches 27.79 / 49.80 / 58.20 at α = 0.1 / 0.5 / 1.0, and LorentzTIM is ahead by 5.71 / 6.00 / 5.30 points. We added α-TIM and tuned TIM under the same Dirichlet protocol (Table D); the sampler is described in our response to Reviewer awPS (W3).

Table D. Imbalanced queries (Dirichlet α, 5-shot), CTX-UXO, DINOv3, macro-F1 (%) ± 95% CI, 1,000 paired episodes. TIM and α-TIM are tuned on validation episodes drawn with the same α; LorentzTIM and TIM (default configuration of Fig. 4b) use the shared default configuration of Fig. 4b (no search); TIM (original protocol) uses TIM's published values. Δ: paired difference with 95% bootstrap CI over episodes.

Methodα = 0.1α = 0.5α = 1.0
SimpleShot31.51 ± 0.7453.15 ± 0.7260.20 ± 0.60
Hyp-SimpleShot31.37 ± 0.7452.80 ± 0.7259.90 ± 0.61
TIM (original protocol, no search)20.86 ± 0.6942.11 ± 0.7851.45 ± 0.71
TIM (default configuration of Fig. 4b, no search)27.79 ± 0.6949.80 ± 0.7858.20 ± 0.71
TIM (tuned)32.39 ± 0.8154.64 ± 0.8261.02 ± 0.69
α-TIM (tuned)29.75 ± 0.6850.08 ± 0.7757.80 ± 0.64
LorentzTIM (default configuration of Fig. 4b, no search)33.50 ± 0.8055.80 ± 0.8063.50 ± 0.69
Δ LorentzTIM − tuned TIM [95% CI]+1.11 [+0.80, +1.42]+1.16 [+0.72, +1.60]+2.48 [+2.10, +2.85]
Δ LorentzTIM − α-TIM [95% CI]+3.75 [+3.35, +4.15]+5.72 [+5.16, +6.28]+5.70 [+5.20, +6.20]

Without any search, LorentzTIM is above tuned TIM at every α (+1.11, +1.16 and +2.48 points at α = 0.1 / 0.5 / 1.0) and above α-TIM by 3.75-5.72 points; all paired intervals exclude zero. α-TIM is searched over its loss weights and its entropy order α ∈ [1.5, 10], the regime for which it was designed (α → 1 recovers TIM, reported in the row above). It falls below tuned TIM because of the tuning: on imbalanced validation episodes the search lowers TIM's marginal-entropy weight to w_ĥ = 0.22 / 0.22 / 0.13 at α = 0.1 / 0.5 / 1.0, against 1.0 in the published TIM, which removes most of the bias towards balanced predictions that α-TIM corrects. With its published hyperparameters, TIM degrades as reported by Veilleux et al. (2021) and falls well below α-TIM (20.86 / 42.11 / 51.45).

Q4. Visual similarity versus class frequency.

We agree that Fig. 4c reports only per-class F1 and measures neither visual similarity nor class frequency. We withdraw the interpretation at L428-431. The correct reading is descriptive: Fig. 4c reports per-class F1 for each method; it does not measure visual similarity or class frequency, and we do not attribute the gains to either.

AuthorsOpenReview ↗

Response to Reviewer g3F3 (4/4)

Minor comments.

We agree with these comments:

  • All F1 values in all responses are on a 0-100 scale (e.g. 53.83 ± 0.94).

  • How to read each diagnostic in Table 2 and the two control participation ratios are explained under Q2 above and in our response to Reviewer pgjA (Q2).

  • We agree that the optimiser details are clearer after the prototype initialisation and the objective.

  • At L324, the objective is Eq. 17 (Eq. 21 with the regulariser); Eq. 19 defines the entropies.

  • We agree that the citations at L69-70 belong at the end of the sentence, and we agree with the suggested positions of Figures 1 and 2.

  • Moreira et al. (2024) are discussed in our response to Reviewer DCaD (W4), and we thank the reviewer for the survey by Mettes, Ghadimi Atigh, Keller-Ressel, Gu and Yeung. This survey will be included in the final manuscript.

Reviewer_awPSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LorentzTIM for few-shot UXO identification. The method keeps the image encoder fixed and performs prototype-based classification and transductive adaptation in Lorentz hyperbolic space. In other words, the paper changes how support and query representations are organized and used for classification, rather than introducing any sort of new image backbone. The geometric diagnostics provide the motivation, and the paper evaluates the approach across four datasets, two feature extractors, and several shot regimes.

Critical strengths
  1. The consistency of the positive empirical results. On CTX-UXO with DINOv3, LorentzTIM shows improvement over the explicitly tuned TIM baseline at every reported shot count. The fixed-parameter LorentzTIM row also exceeds tuned TIM in all four mean macro-F1 results. This is an important control: the observed advantage is not confined to a comparison between a tuned proposal and an untuned baseline.

  2. The appendix clarifies which data fit PCA, which data guide hyperparameter selection, and when the selected settings are fixed. Algorithm 1 fits PCA on support and unlabeled query features, consistent with transductive inference. Algorithm 2 selects hyperparameters using validation-episode macro-D1, and Appendix E states that these choices are fixed before testing. Thus, the PCA fitting pool and validation-only selection procedure are explicitly documented, resolving substantive protocol ambiguities.

  3. The ablation study is informative. At five shots, removing adaptation reduces macro-F1 from 0.721 to 0.671, matching Hyp-SimpleShot in Table 3. The method also includes query-imbalance and extended-shot experiments. Appendix D's advisory framing and approximately 197-199 ms adaptation times in Table 8 are also useful additions.

Critical weaknesses
  1. The interpretation of the geometric control by the authors is stronger than the evidence supports. Normalizing Gaussian samples in Eq. 11 need not preserve their covariance spectrum. I would present the diagnostics as suggestive motivation rather than a demonstration that hierarchy, rather than covariance structure, explains the performance.

  2. The expanded evaluation is not as completely controlled in its reporting as the main DINOv3 experiments. Tables 9-12 do not identify a tuned TIM row, and the selected settings for the other datasets are not given in Tables 6-7. The authors should distinguish retuning from transfer and report the main paired comparisons against tuned TIM. Appendix B does specify that Holm correction is applied across four shot settings for each method; the unresolved issue is the comparator and its uncertainty, not the absence of a statistical procedure.

  3. The query-imbalance experiments could be more informative, limiting the current robustness claim. The marginal-entropy term encourages balanced aggregate predictions, whereas the purpose of the Dirichlet protocol is to vary class proportions. The paper already cites Veilleux et al., whose imbalance-aware information-maximization method is a particularly relevant comparison. Exact imbalance scores, uncertainty intervals, and the conversion from sampled priors to query counts would help establish practical scope.

  4. There are some elements and values in the paper that are not properly explained. Algorithm 2 samples 10-100 steps while Table 5 says 10-1000, and the median-distance scaling of temperature and regularization is not operationally defined. Additionally, Table 2 reports "29.04 / 12.88" for the control's participation ratio without defining the two values. Please clarify which value describes the control used in the geometric diagnostics and how it supports the claimed dimensionality and covariance-spectrum matching.

Questions for authors

1. How should the geometric conclusion be qualified?

Please explain the two control participation-ratio values and state precisely what is preserved by Eq. 11. Can the authors separate the classifier's empirical advantage from the stronger claim about its underlying cause?

2. What tuning protocol was used for the additional tables?

Please identify the TIM settings and whether hyperparameters were retuned or transferred for each dataset/backbone. A direct paired interval against tuned TIM would also strengthen the main comparison.

3. Which implementation details should a reproducer follow?

Please reconcile the two-step count ranges and define the median-distance scaling, including how the values in Tables 6-7 apply at test time.

4. How robust is LorentzTIM when the unlabeled query set contains unequal class proportions, rather than equally many examples of every class?

Please clarify the query-count sampler and compare it with an imbalance-aware information maximization baseline where feasible. This would help the claim made by Figure 4b, in which the method outperforms baselines under severe query imbalance.

4 comments
AuthorsOpenReview ↗

Response to Reviewer awPS (1/4)

We thank the reviewer for the careful assessment and for recognising the fixed-parameter control and the validation-only protocol. We respond to each point below.

Summary.

(1) The geometric control is clarified and complemented by class-conditional and full-covariance controls. '

(2) For CLIP and the three additional datasets, results against both TIM protocols, including a TIM tuned by its own search, with paired tests.

(3) The Dirichlet sampler is specified and α-TIM is added.

AuthorsOpenReview ↗

Response to Reviwer awPS (2/4)

W1. Interpretation of the geometric control.

We agree. We added two stronger controls (Table A), both generated in the full feature space and L2-normalised like Z, without truncation. The UXO embeddings Z have the lowest δ_rel (.2453) and the highest cophenetic correlation (.768 / .613) of the four point sets. The class-conditional control comes close (δ_rel .2678, with overlapping confidence intervals; cophenetic .683 / .567), whereas the two controls without class structure are clearly less hierarchical (δ_rel values are .3070 and .4215, cophenetic at most .520 / .454). The tree-likeness of Z is therefore explained largely by its class structure. As the reviewer recommends, we present the diagnostics as motivation for the hyperbolic model, and the controlled comparisons as the evidence for the method.

Table A. Geometric diagnostics for the UXO embeddings and three controls (DINOv3).

DataParticipation ratioδ_rel (tree-likeness), 95% CIρ_coph (Pearson / Spearman)
UXO embeddings Z29.04.2453 [.213, .274].768 / .613
Single-Gaussian control (Eq. 11)12.88.4215 [.394, .462].520 / .454
Full-covariance Gaussian, no class structure (new)35.29.3070 [.255, .355].505 / .411
Class-conditional Gaussian (new)34.44.2678 [.233, .318].683 / .567

W2. Reporting of the expanded evaluation.

  • Retuning versus transfer. For Tables 9-12, LorentzTIM was re-searched independently for each dataset, backbone and shot setting on that dataset's own validation split, with the same budget as on CTX-UXO (25 trials × 30 validation episodes). No hyperparameters were transferred.

  • TIM comparator. In the submitted Tables 9-12, the TIM row is Euclidean TIM (original implementation). We added a TIM row tuned by its own random search with the same budget (25 trials × 30 validation episodes). Against this tuned TIM, LorentzTIM is ahead in all settings (CTX-UXO with CLIP, and the three additional datasets with both backbones), by +0.86 to +4.47 points, with Holm-adjusted in all 28 settings. All paired differences are listed in our response to Reviewer gF3F (Table B). On the three additional datasets the two TIM rows differ by at most 0.8 points, within their confidence intervals. LorentzTIM exceeds the stronger of the two TIM rows in every setting.

  • Paired comparison with tuned TIM. On CTX-UXO, LorentzTIM exceeds tuned TIM at every shot; the paired statistics are in Table C.

Table C. Paired comparison with tuned TIM, CTX-UXO, DINOv3, 1,000 paired episodes. Δ in percentage points; p-values from a paired Wilcoxon signed-rank test with Holm correction over the four shot settings.

ComparisonStatistic1-shot3-shot5-shot10-shot
LorentzTIM (tuned) vs tuned TIMΔ+4.29+1.92+2.75+4.37
95% CI of Δ[+3.86, +4.72][+1.61, +2.23][+2.41, +3.09][+4.11, +4.63]
Holm-adjusted p< 0.001< 0.001< 0.001< 0.001
LorentzTIM (TIM's fixed hyperparameters) vs tuned TIMΔ+1.74+1.26+2.32+3.43
95% CI of Δ[+1.35, +2.13][+0.95, +1.57][+1.99, +2.65][+3.18, +3.68]
Holm-adjusted p< 0.001< 0.001< 0.001< 0.001

We also evaluated LaplacianShot, PT-MAP and the hyperbolic baselines on the three additional datasets, with paired Wilcoxon-Holm tests; all are below both TIM variants (our response to Reviewer DCaD, Table C).

AuthorsOpenReview ↗

Response to Reviewer awPS (3/4)

W3. Query imbalance and marginal entropy.

We agree that the marginal-entropy term favours balanced aggregate predictions, which works against imbalanced query sets. For each episode we draw class proportions p ~ Dir(α·𝟏₅), convert p to integer query counts over the 75 queries of the balanced protocol by largest-remainder rounding, and keep the support set balanced. Macro-F1 is averaged over all five categories. A category with no queries in an episode contributes F1 = 0 to every method alike (2.56 / 0.57 / 0.12 such classes per episode on average), which lowers the absolute values at small α. Averaging over the classes present instead leaves the ordering of the methods unchanged. Table D reports the imbalance experiment with 95% confidence intervals, includes the more severe setting α = 0.1, and adds α-TIM, which was designed for this setting.

Table D. Imbalanced queries (Dirichlet α, 5-shot), CTX-UXO, DINOv3, macro-F1 (%) ± 95% CI, 1,000 paired episodes. TIM and α-TIM are tuned on validation episodes drawn with the same α; LorentzTIM and TIM (default configuration of Fig. 4b) use the shared default configuration of Fig. 4b (no search). TIM (original protocol) uses TIM's published values. Δ: paired difference with 95% bootstrap CI over episodes.

Methodα = 0.1α = 0.5α = 1.0
SimpleShot31.51 ± 0.7453.15 ± 0.7260.20 ± 0.60
Hyp-SimpleShot31.37 ± 0.7452.80 ± 0.7259.90 ± 0.61
TIM (original protocol, no search)20.86 ± 0.6942.11 ± 0.7851.45 ± 0.71
TIM (default configuration of Fig. 4b, no search)27.79 ± 0.6749.80 ± 0.7558.20 ± 0.72
TIM (tuned)32.39 ± 0.8154.64 ± 0.8261.02 ± 0.69
α-TIM (tuned)29.75 ± 0.6850.08 ± 0.7757.80 ± 0.64
LorentzTIM (default configuration of Fig. 4b, no search)33.50 ± 0.8055.80 ± 0.8063.50 ± 0.69
Δ LorentzTIM − tuned TIM [95% CI]+1.11 [+0.80, +1.42]+1.16 [+0.72, +1.60]+2.48 [+2.10, +2.85]
Δ LorentzTIM − α-TIM [95% CI]+3.75 [+3.35, +4.15]+5.72 [+5.16, +6.28]+5.70 [+5.20, +6.20]

Without any search, LorentzTIM is above tuned TIM at every α (+1.11, +1.16 and +2.48 points at α = 0.1 / 0.5 / 1.0) and above α-TIM by 3.75-5.72 points; all paired intervals exclude zero. α-TIM is searched over its loss weights and its entropy order α ∈ [1.5, 10], the regime for which it was designed (α → 1 recovers TIM, reported in the row above). It falls below tuned TIM because of the tuning: on imbalanced validation episodes the search lowers TIM's marginal-entropy weight to w_ĥ = 0.22 / 0.22 / 0.13 at α = 0.1 / 0.5 / 1.0, against 1.0 in the published TIM, which removes most of the bias towards balanced predictions that α-TIM corrects. With its published hyperparameters, TIM degrades as reported by Veilleux et al. (2021) and falls well below α-TIM (20.86 / 42.11 / 51.45). Table D lists all values with CIs.

W4. Hyperparameter and control reporting.

  • T range. T is sampled from {10, …, 100}, as in Algorithm 2 and the released code (all selected T ≤ 85 on CTX-UXO and T ≤ 95 on the other datasets). The range {10, …, 1000} in Table 5 is a typo.

  • Median-distance scaling. For each shot setting, d̄ is the median, over random pairs (i ≠ j), of the squared geodesic distance d_K(x_i, x_j)² between the lifted support and query points of one validation episode (for TIM, the squared Euclidean distance between the PCA features). τ and ω are then sampled log-uniformly from [10⁻¹, 10^1.5]/d̄ and [10⁻², 10^0.5]/d̄. Tables 6-7 report the selected absolute values, which are used unchanged at test time; d̄ is not recomputed on test episodes.

  • Table 2. 29.04 is the participation ratio of Z and sets d_eff = 29; 12.88 is the participation ratio of the control after truncation and normalisation. Eq. 11 keeps the leading d_eff covariance directions and their variances; the truncation and the normalisation both lower the participation ratio.

AuthorsOpenReview ↗

Response to Reviewer awPS (4/4)

Q1. Qualifying the geometric conclusion.

Yes. The diagnostics are presented as motivation, and the claim follows what the ablations show: hyperbolic geometry gives its main gain in combination with transductive prototype adaptation (Table E). The two participation ratios and what Eq. 11 preserves are explained under W4.

Table E. Differences in macro-F1 (percentage points), CTX-UXO, DINOv3.

Comparison (percentage points)1-shot3-shot5-shot10-shot
Hyp-SimpleShot − SimpleShot (lifting only)+0.69+0.57+0.69+0.41
LorentzTIM − tuned TIM (adaptation on the hyperboloid)+4.29+1.92+2.75+4.37
LorentzTIM with TIM's fixed hyperparameters − tuned TIM+1.74+1.26+2.32+3.43
LorentzTIM with c → 0 (Table 4(a); same regulariser and settings) − tuned TIM−0.68−0.47−0.53−0.56

Accordingly, we withdraw "spectrum-matched" and the stronger claim at L270-272, and keep the narrower claim above.

Q2. Tuning protocol on the additional tables.

LorentzTIM was re-searched independently for each dataset and backbone, with the same budget as on CTX-UXO and no transfer of hyperparameters; Euclidean TIM in the submitted tables reused LorentzTIM's selected hyperparameters, and we added a separately tuned TIM. All paired comparison against tuned TIM is in Table C.

Q3. Implementation details.

The T range is {10, …, 100}, as in Algorithm 2, and the median-distance scaling of τ and ω is defined under W4 above.

Q4. Unequal query proportions.

The Dirichlet sampler is described under W3, and the imbalance experiment is reported with numerical values, 95% confidence intervals and α-TIM as an additional baseline (Table D above).

Reviewer_DCaDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In their work the authors target the problem of unexploded ordnance (UXO) classification in a few-shot setting. They focus on the CTX-UXO dataset and argue that the data shows a level of intrinsic hierarchy that can be better encoded without distortion in hyperbolic space. They show this tree-like structure through multiple tests on the data embeddings based on relative Gromov delta-hyperbolicity, cophenetic correlation, and they argue that the data has an effective dimension of about 29 through a dimension participation ratio. They then use a hybrid hyperbolic model based on a Euclidean DINOv3 encoder, a PCA dimension reducer, and a hyperbolic prototype classifier. The prototypes start at the Lorentzian centroid of the support set and are refined with Riemannian Adam on the TIM objective (support cross-entropy plus query mutual information) with an extra penalty pulling prototypes toward the origin. They show improved performance compared to Euclidean and other hyperbolic baselines.

Critical strengths
  1. The application is typically understudied and benefits from this hyperbolic few-shot approach in identifying the types of ordnance that directs removal methods.
  2. The evaluation protocol seems to be well done with the 1000 fixed-seed episodes shared across all methods, paired Wilcoxon tests with Holm correction, and a TIM baseline given the same random-search budget as the proposed method.
  3. The ablations seem to be well done and isolate performance sources clearly.
Critical weaknesses
  1. Limited methodological novelty where a large portion of the paper is just repeated existing methods including Riemannian optimization explanations, Gromov hyperbolicity... The actual novelty seems to be mainly replacing the Euclidean distance in TIM with the hyperbolic distance metric. The only additions are the origin penalty and the per-shot random search. Swapping the distance is a reasonable thing to try, but on its own it is a small step.
  2. The work focuses on a single benchmark and the appendix experiments on CUB, FGVC-Aircraft and tieredImageNet compare only against SimpleShot and Euclidean TIM. They don't include hyperbolic baselines, LaplacianShot and PT-MAP, and they report no significance tests, and do not say whether TIM was tuned with the same random-search budget as LorentzTIM. Also, tieredImageNet is derived from ImageNet, and DINOv3 and CLIP were could have been pretrained on this subset which could cause some leakage and should be discussed.
  3. The described UXO taxonomy (delivery mechanism -> form factor -> family -> subtype) is used as motivation but is never studied or assessed in the evaluation and no hierarchy-aware metric is reported.
  4. There are some important missing related work and baselines such as Moreira et al. (WACV 2024, "Hyperbolic vs Euclidean Embeddings in Few-Shot Learning: Two Sides of the Same Coin") argue that reported hyperbolic few-shot gains can largely be matched by a fixed-radius Euclidean encoder. This is particularly interesting since it seems here that the embedding radii may be close to the origin as well which is very Euclidean like. Gao et al. (ICCV 2021) is also cited and evaluates hyperbolic few-shot learning in a transductive setting, but is not compared against. And there is an updated alpha-TIM which focuses on imbalanced-query settings, it is also cited but compared against.
Questions for authors
  1. Do the Euclidean baselines also run on the 29 dim PCA features? How do they do on the full 768-d features?
  2. With CL2N, lifted points lie within geodesic radius 1 of the origin. What are the actual radii of points and prototypes?
  3. Were LaplacianShot, PT-MAP, the hyperbolic baselines, and TIM in the appendix tuned with the same search budget?
  4. Section 3 defines labels as subtypes, but Fig. 4c reports five broad categories. What granularity are the episodes sampled at, and how many classes does each data split contain?
9 comments
AuthorsOpenReview ↗

Response to Reviewer DCaD (1/5)

We thank the reviewer for the thorough assessment and for acknowledging the evaluation protocol and the ablation studies. We address each point below.

Summary.

(1) Results on CUB-200-2011, FGVC-Aircraft and tieredImageNet with LaplacianShot, PT-MAP, hyperbolic baselines and a separately tuned TIM, with Wilcoxon-Holm tests.

(2) A hierarchy-aware evaluation with paired confidence intervals is added.

(3) Full-feature, fixed-radius and imbalance controls, and the measured radii are reported.

AuthorsOpenReview ↗

Response to Reviewer DCaD (2/5)

W1. Methodological novelty.

We agree that replacing the Euclidean distance in TIM is the starting point. Beyond the distance, LorentzTIM also defines prototype initialisation, regularisation and optimisation on the Lorentz model (Eqs. 15-25): prototypes start at the Lorentzian centroid and are updated by Riemannian gradients with parallel transport, over a frozen backbone. Two results show that the gain is not explained by the origin penalty or by the per-shot search (Table A):

  • The curvature ablation (the “Curvature c → 0” ablation of Table 4(a)) keeps the origin regulariser and all other LorentzTIM settings, yet scores below tuned TIM at every shot.

  • With TIM's own fixed hyperparameters (no search), LorentzTIM exceeds tuned TIM at every shot, by +1.74 / +1.26 / +2.32 / +3.43 points.

Table A. CTX-UXO, DINOv3, macro-F1 (%), 1,000 paired test episodes.

Method1-shot3-shot5-shot10-shot
TIM, tuned by random search49.5465.1869.3576.14
LorentzTIM, c → 0 (Table 4(a); same regulariser and settings)48.8664.7168.8275.58
LorentzTIM, TIM's fixed hyperparameters (no search)51.2866.4471.6779.57
LorentzTIM, tuned (same budget as TIM)53.8367.1072.1080.51

W2. Cross-dataset baselines and pre-training leakage.

We evaluated LaplacianShot, PT-MAP and the hyperbolic baselines on the three datasets, with paired Wilcoxon signed-rank tests (Holm-corrected) against two TIM variants: TIM run with the hyperparameters selected for LorentzTIM, as in the submitted Tables 10-12, and TIM tuned by its own random search with the same budget. Table B summarises the 1-shot results, and Table C gives all baselines at 1 and 5 shots.

Table B. 1-shot, DINOv3, macro-F1 (%), 1,000 paired episodes per dataset.

MethodCUB-200-2011FGVC-AircrafttieredImageNet
SimpleShot81.4578.3176.12
Euclidean TIM (same ψ as LorentzTIM)83.6680.4678.86
Euclidean TIM (own random search)83.8080.6879.02
LorentzTIM (tuned)86.2483.2082.41
Δ LorentzTIM − TIM (same ψ)+2.58+2.74+3.55
Δ LorentzTIM − TIM (own search)+2.44+2.52+3.39
Holm-adjusted p (Wilcoxon), both comparisons< 0.001< 0.001< 0.001

Table C. Additional baselines, DINOv3, macro-F1 (%), 1,000 paired episodes per setting. Every baseline is below both TIM variants (Holm-adjusted p < 0.001 against tuned TIM).

MethodCUB 1-shotCUB 5-shotFGVC 1-shotFGVC 5-shottiered 1-shottiered 5-shot
SimpleShot81.4590.7678.3189.2776.1287.32
LaplacianShot81.9390.5478.4088.9577.9487.98
PT-MAP82.2890.2678.7688.9177.1687.71
Hyp-ProtoNet81.3290.4478.2489.0476.0386.85
Hyp-Busemann81.5990.6678.3689.2076.2187.43
TIM (same ψ as LorentzTIM)83.6692.1480.4690.7378.8688.99
TIM (own random search)83.8091.9080.6890.4479.0288.65
LorentzTIM86.2493.8283.2092.6582.4191.22

Across all settings, LorentzTIM is above Euclidean TIM, by +1.68 to +3.55 points, and above the separately tuned TIM, by +1.92 to +3.39 points (Holm-adjusted p < 0.001 in every setting, all paired differences are in our response to Reviewer gF3F, Table B).

Tuning protocol. LorentzTIM was searched independently for each dataset, backbone and shot setting on that dataset's validation split, with the same budget as on CTX-UXO. No hyperparameters were transferred from CTX-UXO. Euclidean TIM in the submitted Tables 9-12 reused the hyperparameters selected for LorentzTIM. The added TIM row has its own search.

Pre-training overlap. We agree that tieredImageNet is derived from ImageNet and may overlap with the pre-training data of DINOv3 and CLIP. All methods use the same frozen features and the same paired episodes, so an overlap affects the shared representation rather than any single method. LorentzTIM also improves on CUB-200-2011 and FGVC-Aircraft, which are not derived from ImageNet.

AuthorsOpenReview ↗

Response to Reviewer DCaD (3/5)

W3. The UXO taxonomy is not evaluated.

We agree that the submitted paper did not evaluate the taxonomy. The taxonomy was a domain hypothesis. We now evaluate it directly. The five categories resolve the top two levels of the taxonomy: delivery mechanism and, below it, form factor (here, the launch type). The tree is: Aviation Bomb → air-delivered → free-fall; Mortar Bomb and Projectile → ground-launched → tube-fired; Rocket-Propelled Grenade → ground-launched → rocket-propelled; Grenade → hand-delivered → thrown. The height of the lowest common ancestor (LCA) of two categories is therefore 1 for Mortar Bomb and Projectile, 2 for the other ground-launched pairs and 3 otherwise. For every misclassified query we report (i) whether the predicted category lies in the same delivery group as the true one, and (ii) the mistake severity, i.e. the LCA height of the true and predicted categories (1 to 3, lower is better). Under uniformly random errors, 30% of the errors would stay within the delivery group. To test the same question on a dataset with an official hierarchy, we apply this evaluation to FGVC-Aircraft (manufacturer → family → variant), where the parent group is the aircraft family, and also report the hierarchical distance: the LCA height averaged over all queries, with 0 for correct predictions, so that it reflects both accuracy and the type of error. In both cases the comparison is with TIM tuned by its own random search.

Table D. Hierarchy-aware evaluation of errors, DINOv3, 1,000 paired episodes. Parent group: delivery mechanism (CTX-UXO) and aircraft family (FGVC-Aircraft). Mistake severity: mean LCA height of the true and predicted class over errors (1-3, lower is better).

Dataset, shotsMethodErrors within the true parent group (%)Mistake severity
CTX-UXO, 1-shotTIM (tuned)39.02.492
LorentzTIM42.02.448
CTX-UXO, 5-shotTIM (tuned)39.92.472
LorentzTIM42.42.453
FGVC-Aircraft, 5-shotTIM (tuned)17.92.435
LorentzTIM28.82.199

On CTX-UXO, 39-42% of the errors stay within the true delivery group (chance level: 30%), and LorentzTIM's errors do so more often than those of tuned TIM: +3.0 points [+2.5, +3.5] at 1 shot and +2.5 [+1.8, +3.2] at 5 shots, with a lower mistake severity (−0.044 [−0.050, −0.038] and −0.019 [−0.030, −0.008]). On FGVC-Aircraft, LorentzTIM has a lower hierarchical distance. All intervals are paired 95% bootstrap intervals over episodes and exclude zero.

LorentzTIM therefore not only makes fewer errors. The errors it still makes are more often taxonomically close to the true class.

Reviewer_pgjAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether the geometry of frozen visual embeddings can improve few-shot unexploded ordnance identification. It motivates hyperbolic modeling with geometric diagnostics and adapts TIM to the Lorentz hyperboloid through Lorentzian prototype initialization, geodesic likelihoods, and Riemannian optimization. CTX-UXO results show consistent macro-F1 gains over evaluated Euclidean and hyperbolic baselines, with ablations and additional backbone and dataset evaluations.

Critical strengths
  1. The study addresses a consequential problem with a coherent geometric motivation and a complete adaptation pipeline.
  2. The supplied pseudocode, hyperparameter tables, stated anonymous code release, and human oversight recommendations are further strengths.
Critical weaknesses
  1. Limited Methodological Novelty. The proposed method mainly integrates existing TIM objectives with established hyperbolic operations and Riemannian optimization. The authors should better clarify the methodological novelty and distinguish their approach from related curved-space few-shot learning methods.
  2. Insufficient Geometric Justification. The geometric analysis provides useful motivation, but the connection between the observed tree-like structure and performance improvements remains insufficiently established.
  3. Experimental Fairness and Reproducibility. The improvements over tuned Euclidean TIM require stronger statistical validation. The authors should also clarify the use of query-dependent PCA, inconsistent hyperparameter search ranges, and the computational cost of transductive adaptation under practical deployment settings.
Questions for authors
  1. What methodological capability or insight distinguishes LorentzTIM from TIM combined with established hyperbolic components, and how does it differ from the curved-space adaptation work discussed in Appendix B?
  2. What do the two control participation-ratio values in Table 2 represent, and does the constructed control retain the intended spectrum after truncation and normalization?
  3. Which T range was actually used for the reported random-search results, and what search space was used for the tuned Euclidean TIM comparator?
5 comments
AuthorsOpenReview ↗

Response to Reviewer pgjA (1/3)

We thank the reviewer for the detailed and constructive comments, which we address point by point below.

Summary.

(1) We clarify and substantiate the methodological contribution by identifying what distinguishes LorentzTIM from prior work: Lorentz-native prototype adaptation over frozen Euclidean features, training-free transductive method, supported by controlled experiments showing that the gains arise from hyperbolic transductive adaptation rather than inductive classification or hyperparameter tuning.

(2) We add paired statistical comparisons against tuned TIM (95% CIs, Holm-adjusted p-values).

(3) We clarify the role of Table 2 as motivation only, rather than evidence of the method's superiority.

AuthorsOpenReview ↗

Response to Reviewer pgjA (2/3)

W1. Methodological novelty.

We agree that the individual components (the TIM objective, Lorentzian centroids, Riemannian Adam) are established. What is new is (a) a transductive method whose initialisation, likelihood, optimisation and regularisation are all defined on the Lorentz model over frozen features; (b) controlled evidence on when hyperbolic geometry helps a frozen backbone:

  1. Training-free and defined entirely on the Lorentz model. Prototype initialisation (Lorentzian centroid), the likelihood (geodesic distance), the optimiser (Riemannian Adam with parallel transport) and the regulariser (distance to the origin) all operate on the hyperboloid, over frozen features. Unlike Gao et al. (2021), no component is trained or meta-trained;

  2. Hyperbolic geometry helps mainly through transductive adaptation. Hyp-SimpleShot exceeds SimpleShot at every shot, by +0.41 to +0.69 points, and hyperbolic linear and 2-layer MLP classifiers trained on the support set show no consistent advantage over their Euclidean counterparts (see our response to Reviewer VMSG, Q1), whereas LorentzTIM exceeds tuned Euclidean TIM by +1.92 to +4.37 points. LorentzTIM with T = 0 is Hyp-SimpleShot, so this comparison measures the adaptation alone: at 5 shots it raises macro-F1 from 67.14 to 72.10 (Table 4(b));

  3. The gain is not explained by tuning. With TIM's own fixed hyperparameters, LorentzTIM exceeds TIM tuned by random search at every shot (Table A).

Table A. CTX-UXO, DINOv3, macro-F1 (%), 1,000 paired test episodes.

Method1-shot3-shot5-shot10-shot
TIM, tuned by random search49.5465.1869.3576.14
LorentzTIM, TIM's fixed hyperparameters (no search)51.2866.4471.6779.57
LorentzTIM, tuned (same search budget as TIM)53.8367.1072.1080.51

W2. Geometric justification.

We agree that the diagnostics in Table 2 motivate the hyperbolic model rather than prove that it is needed. The direct evidence comes from controlled comparisons that isolate the contribution of the geometry within the full adaptation procedure (Table B):

  • Hyperbolic lifting alone (SimpleShot → Hyp-SimpleShot) gives a consistent gain of up to +0.69 points in the balanced protocol.
  • Adapting the prototypes on the hyperboloid with the query set (tuned TIM → LorentzTIM) gives +1.92 to +4.37 points.
  • Removing the curvature while keeping every other component, (the “Curvature c → 0” ablation of Table 4(a)), places the method below tuned TIM at every shot.

Table B. Differences in macro-F1 (percentage points), CTX-UXO, DINOv3.

Comparison (percentage points)1-shot3-shot5-shot10-shot
Hyp-SimpleShot vs SimpleShot (lifting only)+0.69+0.57+0.69+0.41
LorentzTIM vs tuned TIM (adaptation on the hyperboloid)+4.29+1.92+2.75+4.37
LorentzTIM (fixed) vs tuned TIM+1.74+1.26+2.32+3.43
LorentzTIM with c → 0 (Table 4(a); same regulariser and settings) vs tuned TIM−0.68−0.47−0.53−0.56

These results support the claim that hyperbolic geometry helps in combination with transductive prototype adaptation. The stronger geometric controls are discussed under Q2.

W3. Fairness, statistics, reproducibility.

We report the following:

  • Same inputs. All methods in Table 3 use the same 29-dimensional PCA representation and the same 1,000 paired test episodes;

  • Same tuning budget. TIM and LorentzTIM are both tuned by random search with 25 trials × 30 validation episodes, over the same ranges for the shared hyperparameters (Q3), and selection uses validation episodes only (Algorithm 2);

  • Statistics against tuned TIM. In the submitted Table 3, the asterisks marked significance against TIM with its original hyperparameters. We now also test against tuned TIM. Paired differences, 95% confidence intervals and Holm-adjusted p-values are given in Table C;

  • Query-dependent PCA. PCA is fitted per episode on the support features and the unlabelled query features. Query labels are never used (Algorithm 1);

  • Cost. Adaptation takes 197-199 ms per 5-way 5-shot episode on one RTX 4090 (Table 8).

Table C. Paired comparison with tuned TIM, CTX-UXO, DINOv3, 1,000 paired episodes. Δ in percentage points. P-values from a paired Wilcoxon signed-rank test with Holm correction over the four shot settings.

ComparisonStatistic1-shot3-shot5-shot10-shot
LorentzTIM (tuned) vs tuned TIMΔ+4.29+1.92+2.75+4.37
95% CI of Δ[+3.86, +4.72][+1.61, +2.23][+2.41, +3.09][+4.11, +4.63]
Holm-adjusted p< 0.001< 0.001< 0.001< 0.001
LorentzTIM (TIM's fixed hyperparameters) vs tuned TIMΔ+1.74+1.26+2.32+3.43
95% CI of Δ[+1.35, +2.13][+0.95, +1.57][+1.99, +2.65][+3.18, +3.68]
Holm-adjusted p< 0.001< 0.001< 0.001< 0.001
AuthorsOpenReview ↗

Response to Reviewer pgjA (3/3)

Q1. Novelty relative to TIM and Gao et al.

Relative to TIM, LorentzTIM keeps the information-maximisation objective but defines every component (initialisation, likelihood, optimisation and regularisation) on the Lorentz model. The gain is not explained by tuning: with TIM's original hyperparameters and no search, LorentzTIM reaches 51.28 / 66.44 / 71.67 / 79.57 macro-F1 at 1 / 3 / 5 / 10 shots, against 49.54 / 65.18 / 69.35 / 76.14 for TIM tuned by random search.

Relative to Gao et al. (2021), the two methods adapt different things and have different training requirements. Gao et al. meta-train an encoder that predicts a task-aware curvature. LorentzTIM requires no training and instead adapts episode-specific prototypes on the hyperboloid over frozen features. At T = 0 it is identical to Hyp-SimpleShot, so Table 4(b) measures the effect of the adaptation.

Q2. Control participation ratios.

The two values measure different quantities, and the control does not retain the spectrum of Z:

  • 29.04 is the participation ratio (PR) of the UXO embeddings Z; it sets d_eff = 29 in Eq. 11.

  • 12.88 is the PR of the Gaussian control after truncation to the leading d_eff directions and L2 normalisation.

Truncating to the leading d_eff directions and normalising to unit norm both change the spectrum of the Gaussian samples, so the control is not spectrum-matched. We introduce two stronger controls: a full-covariance Gaussian (PR 35.29) and a class-conditional Gaussian (PR 34.44), both generated in the full feature space and L2-normalised like Z, without truncation, so their participation ratios are not bounded by d_eff. The UXO embeddings Z have the lowest δ_rel (.2453) and the highest cophenetic correlation (.768 / .613) of the four point sets. The class-conditional control comes close (δ_rel .2678, with overlapping confidence intervals, with cophenetic .683 / .567), whereas the two controls without class structure are clearly less hierarchical (δ_rel values are .3070 and .4215, cophenetic values are .520 / .454). We present the diagnostics as the motivation for the hyperbolic model.

Q3. T range and the tuned-TIM search space.

T is sampled from {10, …, 100}, as in Algorithm 2 and the released code. All selected values are T ≤ 85 (Tables 6-7). The range {10, …, 1000} in Table 5 is a typo. Euclidean TIM was tuned with the same budget (25 trials × 30 validation episodes) and the same ranges for every hyperparameter the two methods share.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to Reviewer pgJA

We sincerely thank the reviewer for pushing us to clarify the distinction between the geometric motivation and the transductive component of our approach. We agree that this distinction is important for positioning the contribution accurately, and we understand that our previous response did not make it sufficiently clear.

The role of geometry. The geometric diagnostics are intended to motivate the use of hyperbolic representations. They do not establish that tree-like structure alone requires a transductive algorithm. We have also qualified this interpretation: the class-conditional control shows that much of the observed tree-likeness is associated with class structure.

The role of transductive adaptation. The transductive formulation is motivated by the deployment setting, in which an unlabelled batch of images is available before a decision is made. The two components are not additive. Hyperbolic geometry amplifies the gain from transduction.

We agree that the individual components are established. Our contribution is not a new optimiser or objective, but their integration into a training-free, Lorentz-native prototype adaptation over frozen features, together with the evidence above on when this combination is beneficial.

We are grateful for the reviewer's time and for the thoughtful questions, which helped us sharpen the positioning of the work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.