acceptodds
Under review as a conference paper at ICLR 2027

VIREO: Learning How to Look and What to Trust for Aerial Object Navigation

Abstract

Aerial object navigation (ObjectNav) requires an unmanned aerial vehicle (UAV) to search for a target specified by a textual description using visual observations. Existing methods mainly rely on spatial search signals to guide navigation, while overlooking that target recognition can vary substantially across observation viewpoints. Thus, approaching a promising candidate does not necessarily ensure reliable recognition. Even at similar distances, the same target can provide substantially different visual evidence from different viewing directions. To bridge this gap, we propose Viewpoint-Informed Recognition and Evidence Optimization (VIREO), a viewpoint-aware framework for aerial ObjectNav. To identify viewpoints that better support target recognition during navigation, VIREO models how recognition outcomes for target and non-target candidates vary across observation viewpoints. Based on these response patterns, VIREO constructs a viewpoint information field to guide the UAV toward viewpoints with higher expected information gain. To address evidence accumulation across viewpoints, VIREO introduces viewpoint-aware evidence aggregation, which adjusts each observation's contribution according to the similarity to previous viewpoints. Experiments on UAV-ON show that VIREO surpasses the best baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
object navigation, unmanned aerial vehicle, vision-language model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_FfXcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VIREO addresses the challenges of viewpoint-dependent recognition and redundant observations in aerial object navigation. It learns viewpoint-conditioned models of VLM responses, selects viewpoints based on expected information gain, and aggregates recognition evidence using geometric novelty weights. Experiments on UAV-ON, supported by online ablations and offline verification analyses, demonstrate promising improvements in overall navigation performance.

Critical strengths
  1. The motivation is supported by empirical observations. Table 3 shows that 54.72% of matching object–description pairs produce mixed recognition responses across viewpoints. This provides concrete support for accounting for viewpoint when verifying candidate objects.

  2. The framework coherently links viewpoint selection with candidate verification. Sections 3.3–3.5 use a shared response model to score prospective observations and update candidate beliefs. The controlled online ablation study in Table 4 supports the usefulness of both components within the same navigation system.

  3. The supplementary evaluation strengthens the empirical contribution. Appendices B–C include verification on held-out scenes, calibration metrics, and bootstrap analyses grouped by object. The paper also explicitly acknowledges the movement-cost trade-off reported in Table 7 and distinguishes the calibration of candidate identity beliefs from navigation success.

Critical weaknesses
  1. Section 3.5 downweights observations from similar camera positions and viewing directions, assuming that they provide redundant evidence. However, the paper does not examine whether similar viewpoints produce correlated recognition errors. Table 3 reports response variation across viewpoints, while Table 6 evaluates final belief calibration. Neither directly tests the relationship between geometric similarity and evidence redundancy.

  2. The response model in Section 3.3 predicts VLM responses using relative camera geometry. It does not explicitly account for candidate appearance, distance to the candidate, or occlusion, all of which can affect a viewpoint’s usefulness. The paper does not report response-prediction accuracy or test whether higher information-gain scores identify more informative views on held-out scenes. Both viewpoint selection and belief updates depend on this model.

  3. In Table 4, enabling evidence aggregation (EA) changes both observation weights and the rules for candidate acceptance and rejection. Consequently, the improvement cannot be attributed specifically to geometric weighting. The policies in Appendix B may also use different observations and stopping times. Controlled comparisons with unit weights and constant weights, using the same response model and decision rules, are needed to isolate the contribution of geometric weighting.

Questions for authors
  1. Do geometrically similar viewpoints produce more strongly correlated recognition errors? Can you provide measurements from the collected observations to support the weighting rule in Section 3.5?

  2. How accurately does the response model predict VLM judgments on held-out scenes? Do viewpoints with higher predicted information gain provide more useful evidence for distinguishing targets from non-targets?

  3. Does geometric weighting outperform unit weighting and constant discounting when the response model and the rules for candidate acceptance and rejection are held fixed?

5 comments
AuthorsOpenReview ↗

The response of weakness_1 and question_1

Comment: We sincerely thank you for your careful review and thoughtful questions. We greatly appreciate your positive assessment of our motivation, unified framework, and supplementary validation, as well as your detailed comments on recognition-error correlation, response-model accuracy, and geometric weighting. Below are our responses.

W1 & Q1: Geometric Similarity and Repeated Errors.

Response: Our measurements show more joint false positives at geometrically similar viewpoints, supporting the use of geometric similarity to identify repeated recognition evidence.

We collect 12 views of 424 objects from 10 training scenes: 5,088 images, 848 matching/non-matching object–description pairs, and 10,176 VLM responses.

For each object–description pair, we rank the 66 possible view pairs by geometric similarity:

Kij=exp⁡(−∥pi−pj∥22σp2−θij22σθ2),K_{ij} = \exp\left(-\frac{\lVert p_i-p_j\rVert^2}{2\sigma_p^2}-\frac{\theta_{ij}^2}{2\sigma_\theta^2}\right),

where pip_i is the camera position, θij\theta_{ij} is the viewing-direction difference, σp=6,m\sigma_p=6\\,\mathrm{m}, and σθ=45∘\sigma_\theta=45^\circ. The 22 highest-scoring and 22 lowest-scoring pairs form the similar and dissimilar groups, respectively.

Using manually annotated visibility labels, we retain pairs with the target absent from both images and measure joint false positives:

J=Pr⁡(yi=1,yj=1∣vi=0,vj=0),J = \Pr(y_i=1,y_j=1\mid v_i=0,v_j=0),

where yi=1y_i=1 denotes a positive VLM response and vi=0v_i=0 denotes annotated target absence. Each sequence must contain at least two eligible pairs in both groups. Rates are averaged within objects and then equally across objects.

We additionally evaluate a restricted subset requiring non-matching descriptions and equal occlusion and known clutter labels within each pair.

The table below reports joint false-positive rates for similar and dissimilar viewpoints:

SubsetObjectsDissimilar (%)Similar (%)
Target absent40414.3630.75
Restricted3916.239.32

Similar views have a 2.14-fold higher joint false-positive rate (30.75% vs. 14.36%); the pattern persists in the restricted subset (9.32% vs. 6.23%), supporting geometric similarity as a cue for repeated recognition evidence.

This measures repeated false positives directly and supports downweighting potentially redundant evidence in Section 3.5.

We have included the view-pair analysis in Appendix A.3, “Geometric Similarity and Evidence Redundancy” (Table 7) of the revised manuscript.

AuthorsOpenReview ↗

The response of weakness_2 and question_2

W2 & Q2: Response Prediction and Viewpoint Utility.

Response: Our held-out-scene results show improved response prediction for true candidates, while the full viewpoint score improves true-candidate acceptance and observation efficiency in saved-observation verification.

(a) Response prediction.

We compare a constant-response model with the geometry-conditioned model over five scene-disjoint folds, keeping training, parameter-selection, and test scenes separate. Each candidate type contains 5,088 observations. The models predict the VLM's binary response, evaluated using NLL, Brier score, and ECE. For response yy and predicted positive-response probability pp, NLL averages −[ylog⁡p+(1−y)log⁡(1−p)]-[y\log p+(1-y)\log(1-p)], and Brier averages (p−y)2(p-y)^2. ECE measures the frequency-weighted calibration gap over ten equal-width probability bins. Lower scores are better.

Constant predicts a fixed positive-response probability for each candidate type, independent of viewpoint.

Geometry predicts this probability from viewpoint geometry. Both models predict VLM responses.

The table below reports VLM response prediction on held-out scenes:

CandidateModelNLL ↓Brier ↓ECE ↓
MatchingConstant0.614050.209310.09943
MatchingGeometry0.608850.206200.07254
Non-matchingConstant0.336480.084570.06263
Non-matchingGeometry0.301260.074650.04683

For both matching and non-matching candidates, conditioning on viewpoint geometry yields lower NLL, Brier score, and ECE than fixed-probability prediction, indicating more accurate and better-calibrated predictions of VLM responses.

(b) Viewpoint selection.

We replay saved observations for 424 objects and 848 matching/non-matching candidates. All selectors share the response model, evidence weights, initial view, initial belief of 0.5, decision thresholds of 0.95/0.05, and a six-observation budget including the first image. We vary only the next-view selector: predicted information gain multiplied by geometric novelty, novelty alone, or uniform random selection. Random selection is averaged over 64 repeats per candidate before object-level aggregation.

The full score combines predicted mutual information with novelty 1−max⁡jKij1-\max_j K_{ij}. For current belief bb and predicted positive-response probabilities r1r_1 and r0r_0 under target and non-target identities, respectively,

MI=H(br1+(1−b)r0)−bH(r1)−(1−b)H(r0),\mathrm{MI} = H(br_1+(1-b)r_0)-bH(r_1)-(1-b)H(r_0),

where H(t)=−tlog⁡t−(1−t)log⁡(1−t)H(t)=-t\log t-(1-t)\log(1-t).

TAR and FAR are the fractions of true and non-target candidates accepted, respectively. The undecided rate covers candidates without a decision after six observations; mean observations counts views until a decision or budget exhaustion.

The table below reports candidate verification under alternative view selectors:

SelectorTAR (%) ↑FAR (%) ↓Undecided (%) ↓Mean views ↓
MI × novelty79.016.607.192.71
Novelty only75.947.088.252.77
Random65.9011.6110.173.28

Compared with random selection and novelty-only selection, the full score increases TAR by 13.11 and 3.07 percentage points, respectively. It also yields lower FAR and undecided rates, while requiring approximately 17.4% fewer observations than random selection. These results support the effectiveness and efficiency of combining predicted information gain with geometric novelty for candidate verification.

We have included the held-out-scene response prediction and viewpoint-selection comparison in Appendix A.5, “Response Prediction and Viewpoint Selection” (Tables 13–14) of the revised manuscript.

AuthorsOpenReview ↗

The response of weakness_3 and question_3

W3 & Q3: Independent Contribution of Geometric Weighting.

Response: Geometric novelty weighting achieves higher SR, OSR, and SPL than unit weighting, fixed discounting, and visual-content-based weighting in the controlled online comparison.

We replace only the evidence-weighting rule, fixing the response model, viewpoint planner, candidate acceptance/rejection rules, and other components. All methods use Qwen3-VL-plus, RGB observations, official target descriptions, the same 200-task subset, a 150-action limit, and the same API-call rules per planning step.

The table below summarizes the alternative evidence-weighting rules:

WeightWeighting rule
UnitAll observations retain a weight of 1
Fixed discountingEach new observation starts at 1; earlier evidence weights decay by a factor of 0.5 per subsequent observation
RGBNovelty based on RGB pixel similarity between the current and previous images
ResNet18Novelty based on frozen ResNet18 feature similarity between the current and previous images

We evaluate SR, OSR, and SPL through closed-loop navigation. The table below reports online navigation performance under alternative weighting rules (%):

WeightSR ↑OSR ↑SPL ↑
Geometry25.5037.0017.41
Unit17.0032.5012.27
Fixed discounting13.5029.5010.72
RGB18.5033.0011.89
ResNet1813.5034.508.66

The higher navigation metrics support the contribution of geometric weighting with the response model and candidate-decision rules held fixed.

We have included the weighting rules and online comparison in Section 4.3, under “Novelty Weighting Ablation” (Eqs. 15–16 and Table 5) of the revised manuscript.

We hope these results clarify your concerns. Thank you again for your thoughtful feedback.

Reviewer_h3mWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VIREO addresses conflicting recognition across viewpoints in aerial object navigation by coupling active observation with evidence accumulation. An offline viewpoint-conditioned response model predicts information gain, while geometric novelty weights recognition evidence for candidate belief and navigation decisions. Across UAV-ON's test environments, the paper reports 24.56% SR and 16.59% SPL, with online ablations and offline cross-scene replay evaluating its components.

Critical strengths
  1. Viewpoint selection and belief updates share a response model, explicitly linking observation acquisition with evidence assessment.
  2. Four online variants share 200 tasks, perception, and action budgets. The full system improves SR by 12 percentage points over the base agent, supporting the combined design.
Critical weaknesses
  1. Calibration uses balanced offline queries, whereas online candidates are selected through positive responses and retain a fixed 0.5 prior.
  2. The response model uses eight geometric features relative to the first viewpoint and shares parameters across all target descriptions, so the information field cannot adapt to target appearance.
  3. Stopping depends only on identity belief reaching 0.95 without a distance check, while success requires stopping within 20 m. More reliable long-range recognition therefore makes stops outside the success radius more likely.
Questions for authors
  1. The authors should report the fraction of activated online candidates that are true targets and the empirical target rate for candidates whose belief falls between 0.9 and 1.0, to show whether ρinit = 0.5 suits candidates selected by positive responses.
2 comments
AuthorsOpenReview ↗

The response of weakness_ 1, question_1 and weakness_3

Comment: We greatly appreciate your recognition of the shared response model linking viewpoint selection and belief updates, together with our controlled online comparisons. Thank you for closely examining candidate initialization, description dependence, and stopping behavior. We provide our responses below.

W1 & Q1: Candidate Prevalence and Prior Sensitivity.

Response: Our results show a higher true-target proportion among high-belief candidates and tolerance to a range of initial ρinit\rho_{\mathrm{init}}. We use 0.5 as a neutral initialization, rather than the target frequency after positive-response selection.

(a) Online candidate measurements.

We measure true-target proportions among first activated candidates and those with beliefs between 0.9 and 1.0.

The table below reports mean online candidate counts and true-target proportions:

CategoryMean candidatesMean true targetsProportion
First activated online candidates405.1311.176.79%
Belief falls between 0.9 and 1.0295287.397.38%

The higher true-target proportion at high belief supports the role of evidence accumulation in candidate verification.

(b) Initial-belief sensitivity.

On the same 200 tasks, we vary only the initial belief across the seven settings below, including the original 0.5. Lower priors delay acceptance; higher priors encourage earlier acceptance. We exclude 0 and 1 because subsequent evidence cannot update these extreme beliefs.

The table below reports navigation performance with different initial candidate beliefs (%):

ρinit\rho_{\mathrm{init}}SROSRSPL
0.1016.5025.008.56
0.2520.5029.0013.81
0.4024.0034.5016.28
0.5025.5037.0017.41
0.6024.5036.5018.51
0.7519.5029.5014.17
0.9013.0019.5011.61

These results suggest that VIREO is relatively insensitive to variations in ρinit\rho_{\mathrm{init}} within a range around the original setting.

We have included the online candidate measurements in Appendix A.7 (Table 17) and the initial-belief ablation in Section 4.3, under “Initial Candidate Belief Ablation” (Table 6) of the revised manuscript.

W3: Stopping Criteria and Success Radius.

Response: Our analysis suggests that VIREO’s evidence-based stopping remains effective without an explicit distance threshold. The observed out-of-radius stops are partly attributable to UAV-ON’s center-based distance criterion for large targets, which can classify stops close to the object surface as failures because the UAV remains outside the prescribed radius from the object’s center.

(a) Stopping design.

Explicit distance estimation requires target localization and segmentation, which are affected by viewpoint-dependent recognition (paper Table 3) and missing or incorrect depth in some UAV-ON scenes. VIREO instead uses accumulated identity evidence. Training viewpoints are sampled with the 20 m target region as the final destination, linking calibration to near-target observations. Without an explicit distance check, the confidence threshold can encourage further approach when distant evidence is insufficient.

(b) Stopping-distance analysis.

We group outcomes by object size and measure the proportion of episodes that track the true target but stop beyond 20 m.

The table below reports stops beyond the success radius by object size:

Object sizeStops beyond 20 m (%)
Small9.38
Mid6.72
Large21.49

We consider the proportions for small and medium targets acceptable for the current design. The higher proportion for large targets appears partly related to center-based distance evaluation: stopping close to a large object's surface can still leave the UAV beyond the 20 m radius from its center.

We have included the stopping design and target-size analysis in Appendix A.7, “Online Candidate Beliefs and Stopping Behavior” (Table 18) of the revised manuscript.

AuthorsOpenReview ↗

The response of weakness_2

W2: Description Sensitivity.

Response: Adding text improves held-out response-probability estimates, while selected-view recognition accuracy remains close to that of the geometry-only model.

UAV-ON provides a fixed appearance description for each target. We retain these descriptions for navigation evaluation, as changing them could alter task difficulty and affect comparability with benchmark results. To examine description sensitivity, we conduct a separate offline evaluation using additional descriptions emphasizing category, appearance, and detail. By keeping objects and images fixed while varying descriptions, we assess how text conditioning affects VLM response prediction and recognition accuracy at selected viewpoints.

(a) Response prediction.

We fix objects and images while varying model-assisted descriptions of category, appearance, and detail, each with matching/non-matching versions. For 193 objects with 12 views each, this yields 2,316 images, 1,158 object–description pairs, and 13,896 VLM queries. We compare three response predictors.

The table below summarizes the response models used in the description-sensitive evaluation:

ModelPredicted response probabilityInputs
Geometry onlyp=σ(b+wg⊤g)p=\sigma(b+w_g^\top g)Viewpoint geometry
Geometry + textp=σ(b+wg⊤g+wt⊤t)p=\sigma(b+w_g^\top g+w_t^\top t)Additive geometry and text features
Interactionsp=σ(b+wg⊤g+wt⊤t+g⊤Wt)p=\sigma(b+w_g^\top g+w_t^\top t+g^\top Wt)Geometry, text, and their interactions

Here, gg and tt are geometry and text features, and σ\sigma is the sigmoid. Interactions allow geometric effects to vary with text. We use shared scene-disjoint splits, fit text features on training data only, and evaluate response probabilities with NLL and Brier score:

NLL=−1M∑k=1M[yklog⁡pk+(1−yk)log⁡(1−pk)],\mathrm{NLL} = -\frac{1}{M}\sum_{k=1}^{M}\left[y_k\log p_k+(1-y_k)\log(1-p_k)\right],

Brier=1M∑k=1M(pk−yk)2,\mathrm{Brier} = \frac{1}{M}\sum_{k=1}^{M}(p_k-y_k)^2,

Here, yky_k is the recorded VLM response and pkp_k its predicted positive-response probability; lower scores are better.

The table below reports response prediction on held-out scenes:

ModelNLL ↓Brier ↓
Geometry only0.57550.1954
Geometry + text0.56930.1919
Interactions0.57090.1923

Adding text reduces NLL by 1.07% and Brier score by 1.76%.

(b) Selected-view accuracy.

From the same reference image, each model selects the highest-information-gain view among the remaining 11. We score its recorded Qwen response as correct if positive for a matching candidate or negative otherwise, averaging over 193×3×2=1,158193\times3\times2=1,158 pairs.

The table below reports recognition accuracy at selected viewpoints:

SelectorAccuracy (%)
Geometry only74.09
Geometry + text74.18
Interactions73.06

Geometry + text reaches 74.18%, a 0.09-percentage-point gain, while interactions reach 73.06%. The similar accuracy of the two main models suggests that text primarily improves response-probability estimation here.

We have included the response-model definitions, description-sensitive response prediction, and selected-view recognition results in Appendix A.4, “Description Sensitivity” (Tables 10–12) of the revised manuscript.

We hope these results clarify your concerns. Thank you again for your thoughtful feedback.

Reviewer_YB9KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents VIREO, a framework for aerial object navigation that combines informative viewpoint selection with viewpoint-aware evidence aggregation. It learns viewpoint-conditioned VLM response models to estimate information gain and weights accumulated recognition evidence according to geometric novelty. Experiments on UAV-ON, together with online ablations and offline verification analyses, show improvements in navigation performance and candidate-belief calibration.

Critical strengths
  1. The motivation is clear and practically relevant: approaching a candidate does not guarantee reliable recognition, and repeated observations may provide redundant evidence.
  2. The two components form a coherent framework, connecting viewpoint planning with evidence accumulation and candidate decisions.
  3. The evaluation includes navigation comparisons, component ablations, held-out-scene verification, and calibration analyses with bootstrap uncertainty estimates.
Critical weaknesses
  1. Geometric similarity may not reflect evidence redundancy under clutter, occlusion, or large viewpoint changes. Similar poses can reveal different target details, while distant viewpoints can provide redundant or irrelevant observations. Robustness under clutter, occlusion, and large viewpoint changes remains insufficiently evaluated.

  2. The EA ablation changes both weighting and decision rules, leaving novelty weighting’s independent contribution unclear. The online EA ablation changes both evidence weighting and candidate-decision rules. Although offline policy comparisons are provided, they do not fully isolate the proposed weight from other policy differences.

  3. The response model lacks explicit text conditioning, limiting its ability to select description-specific viewpoints. Response models pool observations across descriptions and use geometric features, although informative viewpoints may depend on the particular attributes requested by the text. The method’s ability to support fine-grained, description-dependent viewpoint selection is therefore unclear.

  4. Baseline VLMs and inference budgets are not clearly matched, weakening attribution of the reported gains. Strong-baseline results are taken from prior publications without establishing matched VLM and computational settings. The controlled online ablation uses only 200 tasks, limiting conclusions about the source and consistency of the overall gains.

Questions for authors
  1. How well does geometric similarity predict actual recognition redundancy? Can you provide targeted results under occlusion, clutter, and large viewpoint changes, including candidate correspondence failures?

  2. Could you isolate novelty weighting by keeping the response model, viewpoint planner, and candidate-decision rules fixed, and compare it against unit weights, fixed discounting, and visual-content-based weighting?

  3. For the same object, different descriptions may require observing different attributes. How does the geometry-only response model accommodate this? Can you provide a description-sensitive evaluation or a text-conditioned alternative?

  4. Can you clarify whether the main baselines use comparable VLMs, observation inputs, and inference budgets, and report variability and VLM-call costs for the controlled comparisons?

  5. Could you better position VIREO within active multi-view recognition and view-disentangled representation learning? For example, Multi-View Active Fine-Grained Visual Recognition (ICCV 2023) studies active view selection and stopping, while DRDN Dual-Recommendation Disentanglement Network for View Fuzz in Action Recognition (TIP 2023) addresses viewpoint-induced interference. How does VIREO differ from these approaches, and what evidence supports geometric novelty as a proxy for complementary recognition information?

Then see the weaknesses. A clear and convincing response to these concerns could lead me to reconsider my final recommendation.

6 comments
AuthorsOpenReview ↗

The response of weakness_1 and question_1

Comment: We sincerely thank you for recognizing VIREO’s practical motivation, coherent integration of viewpoint planning and evidence aggregation, and supporting evaluation. We appreciate your detailed questions about geometric novelty, description sensitivity, and comparison fairness. Below, we address your concerns point by point.

W1 & Q1: Geometric Redundancy and Robustness.

Response: Our results support geometric similarity as a cue for repeated recognition evidence, with lower candidate NLL under geometric weighting in the tested challenging subsets.

(a) Repeated false positives.

We collect 12 views of 424 objects from 10 training scenes: 5,088 images, 848 matching/non-matching object–description pairs, and 10,176 VLM responses.

For each object–description pair, we rank the 66 possible view pairs by geometric similarity:

Kij=exp⁡(−∥pi−pj∥22σp2−θij22σθ2),K_{ij} = \exp\left(-\frac{\lVert p_i-p_j\rVert^2}{2\sigma_p^2}-\frac{\theta_{ij}^2}{2\sigma_\theta^2}\right),

where pip_i is the camera position, θij\theta_{ij} is the viewing-direction difference, σp=6,m\sigma_p=6\\,\mathrm{m}, and σθ=45∘\sigma_\theta=45^\circ. The 22 highest-scoring and 22 lowest-scoring pairs form the similar and dissimilar groups, respectively.

Using manually annotated visibility labels, we retain pairs with the target absent from both images and measure joint false positives:

J=Pr⁡(yi=1,yj=1∣vi=0,vj=0),J = \Pr(y_i=1,y_j=1\mid v_i=0,v_j=0),

where yi=1y_i=1 denotes a positive VLM response and vi=0v_i=0 denotes annotated target absence. Each sequence must contain at least two eligible pairs in both groups. Rates are averaged within objects and then equally across objects.

We additionally evaluate a restricted subset requiring non-matching descriptions and equal occlusion and known clutter labels within each pair.

The table below reports joint false-positive rates for similar and dissimilar viewpoints:

SubsetObjectsDissimilar (%)Similar (%)
Target absent40414.3630.75
Restricted3916.239.32

Similar views have a 2.14-fold higher joint false-positive rate (30.75% vs. 14.36%); the pattern persists in the restricted subset (9.32% vs. 6.23%), supporting geometric similarity as a cue for repeated recognition evidence.

(b) Occlusion and clutter.

We compare geometric weighting with constant discounting, fixing observations, VLM responses, view order, response model, initial belief (0.5), and decision thresholds (0.95/0.05). Weighting parameters are selected on separate calibration scenes.

Subsets require occlusion or clutter in at least one of either the first six views or the full 12-view set. Both protocols evaluate beliefs after the same first six observations using object-balanced candidate NLL:

NLL=−1O∑o=1O1∣So∣∑s∈So[ξslog⁡ρs(6)+(1−ξs)log⁡(1−ρs(6))],\mathrm{NLL} = -\frac{1}{O}\sum_{o=1}^{O}\frac{1}{|S_o|}\sum_{s\in S_o}\left[\xi_s\log\rho_s^{(6)}+(1-\xi_s)\log\left(1-\rho_s^{(6)}\right)\right],

where SoS_o contains the eligible sequences for object oo, ξs\xi_s is the binary candidate-match label, and ρs(6)\rho_s^{(6)} is the estimated matching probability after six observations. Lower NLL indicates better probability estimates for candidate identity.

The table below reports six-view candidate NLL under occlusion and clutter:

Subset selectionConditionObjectsSequencesConstant ↓Geometry ↓
First six viewsHigh clutter931861.0245130.823829
First six viewsPartial occlusion34340.2131260.211713
Full 12-view setHigh clutter1362721.0534250.765194
Full 12-view setPartial occlusion50510.3608050.346103

Geometric weighting lowers NLL in all four subsets: by 19.59% and 27.36% under clutter, and 0.66% and 4.07% under partial occlusion, for the two selection protocols respectively. This supports its usefulness for candidate-belief estimation, particularly under clutter.

(c) Candidate correspondence failures.

We construct correspondence-failure sequences containing different objects (468 observations; 936 evidence updates). With the response model, observation order, and decision rules fixed, we compare weighting methods using undecided rate and candidate NLL:

The table below reports candidate verification with correspondence failures:

Weighting methodUndecided rate (%) ↓NLL ↓
Constant29.490.5446
Geometry12.820.4780

Geometric weighting reduces the undecided rate by 56.52% and NLL by 12.23%, supporting candidate verification in these correspondence-failure sequences.

We have included these analyses in Appendix A.3, “Geometric Similarity and Evidence Redundancy” (Tables 7–9) of the revised manuscript.

AuthorsOpenReview ↗

The response of weakness_ 2, question_2, weakness_ 3 and question_3

W2 & Q2: Contribution of Novelty Weighting.

Response: Our online results show that geometric novelty weighting achieves higher SR, OSR, and SPL than the alternative weighting schemes.

We replace only the weighting rule, fixing the response model, viewpoint planner, candidate-decision rules, and all other components:

The table below reports alternative evidence-weighting rules:

WeightWeighting rule
UnitAll observations retain a weight of 1
Fixed discountingEach new observation starts at 1; earlier evidence weights decay by a factor of 0.5 per subsequent observation
RGBNovelty based on RGB pixel similarity between the current and previous images
ResNet18Novelty based on frozen ResNet18 feature similarity between the current and previous images

We compare SR, OSR, and SPL through closed-loop navigation on the paper's same 200-task subset.

The table below reports online navigation under alternative weighting rules (%):

WeightSR ↑OSR ↑SPL ↑
Geometry25.5037.0017.41
Unit17.0032.5012.27
Fixed discounting13.5029.5010.72
RGB18.5033.0011.89
ResNet1813.5034.508.66

These results support the independent contribution of geometric novelty weighting to navigation performance. We have included the weighting rules and online comparison in Section 4.3(Novelty Weighting Ablation) of the revised manuscript.

W3 & Q3: Description Sensitivity.

Response: Adding text improves held-out response-probability estimates, while selected-view recognition accuracy remains close to that of the geometry-only model.

UAV-ON provides a fixed appearance description for each target. We retain these descriptions for navigation evaluation, as changing them could alter task difficulty and affect comparability with benchmark results. To examine description sensitivity, we conduct a separate offline evaluation using additional descriptions emphasizing category, appearance, and detail. By keeping objects and images fixed while varying descriptions, we assess how text conditioning affects VLM response prediction and recognition accuracy at selected viewpoints.

(a) Response prediction.

We fix objects and images while varying model-assisted descriptions of category, appearance, and detail, each with matching/non-matching versions. For 193 objects with 12 views each, this yields 2,316 images, 1,158 object–description pairs, and 13,896 VLM queries. We compare three response predictors:

Response models used in the description-sensitive evaluation.

ModelPredicted response probabilityInputs
Geometry onlyp=σ(b+wg⊤g)p=\sigma(b+w_g^\top g)Viewpoint geometry
Geometry + textp=σ(b+wg⊤g+wt⊤t)p=\sigma(b+w_g^\top g+w_t^\top t)Additive geometry and text features
Interactionsp=σ(b+wg⊤g+wt⊤t+g⊤Wt)p=\sigma(b+w_g^\top g+w_t^\top t+g^\top Wt)Geometry, text, and their interactions

Here, gg and tt are geometry and text features, and σ\sigma is the sigmoid. Interactions allow geometric effects to vary with text. We use shared scene-disjoint splits, fit text features on training data only, and evaluate response probabilities with NLL and Brier score:

NLL=−1M∑k=1M[yklog⁡pk+(1−yk)log⁡(1−pk)],\mathrm{NLL} = -\frac{1}{M}\sum_{k=1}^{M}\left[y_k\log p_k+(1-y_k)\log(1-p_k)\right],

Brier=1M∑k=1M(pk−yk)2,\mathrm{Brier} = \frac{1}{M}\sum_{k=1}^{M}(p_k-y_k)^2,

Here, yky_k is the recorded VLM response and pkp_k its predicted positive-response probability; lower scores are better.

Held-out-scene response prediction.

ModelNLL ↓Brier ↓
Geometry only0.57550.1954
Geometry + text0.56930.1919
Interactions0.57090.1923

Adding text reduces NLL by 1.07% and Brier score by 1.76%.

(b) Selected-view accuracy.

From the same reference image, each model selects the highest-information-gain view among the remaining 11. We score its recorded Qwen response as correct if positive for a matching candidate or negative otherwise, averaging over 193×3×2=1,158193\times3\times2=1,158 pairs.

Recognition accuracy at selected viewpoints.

SelectorAccuracy (%)
Geometry only74.09
Geometry + text74.18
Interactions73.06

Geometry + text reaches 74.18%, a 0.09-percentage-point gain over geometry only. The similar accuracy suggests a small effect on selected-view responses. We retain VIREO's original navigation configuration and parameters for UAV-ON's prescribed descriptions.

We have included the response-model definitions, description-sensitive response prediction, and selected-view recognition results in Appendix A.4, “Description Sensitivity” (Tables 10–12) of the revised manuscript.

AuthorsOpenReview ↗

The response of weakness_ 4, question_4 and question_5

W4 & Q4: Comparison Fairness, Variability, and Cost.

Response: The experiments we conducted share same VLM, observation inputs, and action budget.

Our audit of execution logs and API records covers the main-result evaluation and all 200-task ablations. These experiments use Qwen3-VL-plus, RGB inputs, official UAV-ON descriptions, a limit of 150 physical actions per episode, and equal API calls per planning step.

(a) Performance variability.

The table below reports navigation performance on the full test set and the 200-task subset, with the corresponding ranges in brackets:

MethodSROSRSPL
VIREO-full24.56 [20.87, 27.01]38.06 [33.83, 40.93]16.59 [14.87, 19.92]
VIREO-20025.50 [20.00, 31.50]37.00 [30.50, 43.50]17.41 [12.88, 22.14]
Unit17.00 [12.00, 22.00]32.50 [26.50, 39.00]12.27 [8.36, 16.47]
Fixed discount13.50 [9.00, 18.00]29.50 [23.50, 35.50]10.72 [6.99, 14.73]
RGB18.50 [12.00, 22.00]33.00 [27.00, 39.00]11.89 [8.06, 16.03]
ResNet1813.50 [9.50, 18.00]34.50 [28.50, 40.50]8.66 [5.56, 12.05]

(b) Inference cost.

On the same 200 tasks, we report total logical VLM calls, mean calls/task with the corresponding ranges, and mean recorded tokens/task.

The table below reports inference cost on the 200-task subset:

MethodTotal callsCalls/taskTokens/task
VIREO9,68748.44 [43.94, 53.00]183,550
Unit7,91239.56 [35.30, 43.88]154,298
Fixed discount7,38136.91 [32.70, 41.17]143,755
RGB7,86739.34 [35.12, 43.61]151,976
ResNet187,77338.87 [34.57, 43.15]151,323

Lower call counts with alternative weights can reflect premature termination or incorrect stopping; realized cost should therefore be considered alongside navigation performance.

We have included the shared evaluation settings, performance ranges, and inference costs in Appendix A.6, “Navigation Variability and Inference Cost” (Tables 15–16) of the revised manuscript.

Q5: Relation to Prior Work.

Response: VIREO's focus is the integration of viewpoint planning and viewpoint-aware evidence aggregation during aerial ObjectNav.

We have carefully read both papers you mentioned and appreciate the opportunity to clarify how their objectives differ from ours. MAFR studies active view selection and dynamic exiting for fine-grained recognition, while DRDN disentangles action and view representations to reduce viewpoint interference in action recognition. VIREO focuses on joint viewpoint planning and viewpoint-aware evidence aggregation in aerial object navigation tasks.

Our view-pair analysis shows more joint false positives for similar views (Appendix A.3, Table 7), while the controlled weighting comparison shows higher SR, OSR, and SPL with geometric weighting (Section 4.3, Table 5). These results support geometric novelty as a proxy for complementary evidence: weighting observations by camera position and viewing direction helps reduce the contribution of repeated evidence, rather than relying solely on temporal order.

We hope these results clarify your concerns. Thank you again for your thoughtful feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.