acceptodds
Under review as a conference paper at ICLR 2027

Robust Spatio-Temporal Adaptation towards 4D Point Cloud Understanding

Abstract

4D point cloud understanding jointly models fine-grained 3D geometry and temporal dynamics, playing a crucial role in robotics, autonomous driving and AR/VR. However, point cloud videos are often degraded by hardware noise, object occlusion and stream disruptions in real-world deployment. We reveal that existing state-of-the-art 4D point cloud understanding methods are vulnerable to such disturbances and formulate a unified evaluation protocol for spatial and temporal corruptions. To address these intrinsic weaknesses, we further propose a novel training-free method, Spatio-TEmporal Adaptation with Decoding dYnamics (STEADY), which adapts corrupted geometric evidence and recovers disrupted temporal dynamics at test time. Specifically, our proposed STEADY method comprises two stages: (1) The Hierarchical Spatial Cache retrieves complementary global and local geometric cues while Soft Temporal Consistency Gate modulates cache corrections based on inter-frame consistency. (2) The Dynamics-Aware Decoder incorporates dynamics of transition and duration priors to recover temporally coherent predictions over the complete sequence. We conduct extensive experiments on three widely used 4D point cloud understanding benchmarks, including MSR-Action3D, NTU-RGBD60 and HOI4D, under eight corruption conditions across both spatial and temporal dimensions. Experimental results demonstrate that STEADY consistently improves the overall robustness of various representative backbones across spatial and temporal corruptions. Our anonymous code is available at https://anonymous.4open.science/r/STEADY/.

Primary area:
learning on time series and dynamical systems
Keywords:
Point Cloud Video Understanding, Corruption Robustness, Test-Time Adaptation

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_ng82OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies corruption robustness in 4D point cloud understanding, covering action recognition and temporal action segmentation. It introduces an evaluation protocol with five spatial corruptions and three temporal perturbations on MSR-Action3D, NTU-RGBD60, and HOI4D, and evaluates several representative backbones. It also proposes STEADY, which combines an online hierarchical cache of global and local point-tube features with temporal consistency gating. For action segmentation, a dynamics-aware decoder further combines the adapted frame scores with initial-state, transition, and duration priors estimated from training annotations through explicit-duration hidden semi-Markov decoding. Backbone weights remain frozen.

The experiments show modest improvements in average corrupted recognition accuracy and substantially larger improvements in average segmentation edit and F1 scores. However, frame accuracy decreases for some segmentation backbones, as acknowledged in the appendix. Ablations, sensitivity analyses, qualitative examples, and inference-cost measurements provide additional validation.

Critical strengths
  • The paper investigates a relevant and comparatively underexplored robustness problem with useful experimental breadth: eight corruption types, three datasets, two tasks, and six distinct backbone architectures across the experiments. Reporting results by corruption type exposes differences that clean performance alone does not reveal. A sufficiently specified and released protocol could provide a useful basis for subsequent work on point cloud video robustness.
  • STEADY achieves substantial average segment-level improvements across all four evaluated segmentation backbones without updating neural-network weights. For example, average edit score increases from 60.1 to 73.7 for P4Transformer and from 80.9 to 86.2 for Mamba4D; average [email protected] also improves for both models and the other segmentation backbones. These improvements provide meaningful evidence that the proposed inference procedure can produce more coherent action segments under corruption (Tables 2 and 6).
Critical weaknesses
  • The comparisons do not establish how much HSC and STCG contribute beyond applying the decoder directly to the backbone outputs. A backbone-plus-DAD-only comparison is missing. In Table 3, HSC and STCG together improve spatial and temporal edit scores by 0.41 and 1.12 points, respectively, whereas subsequently adding DAD improves them by another 3.14 and 7.02 points. These results show that the decoder is important within the tested combination, but do not establish the additional benefit of the cache and gate over decoding the original predictions.

  • The paper does not clearly show whether its cache design performs better than existing cache-based methods. Figure 3 tests versions without local features or negative cache entries, but it is unclear whether these versions are adaptations of TDA or Point-Cache. Explaining this connection and comparing results using the same backbone and evaluation settings would help show the benefit of the proposed design.

  • The data and criterion used for hyperparameter selection are unclear, as is the ordering of the recognition test stream. These details matter because the method uses several cache and decoding parameters, and the recognition cache is updated sequentially across test observations. A clearer account of parameter selection and stream construction is needed to assess the evaluation protocol.

  • The paper does not clearly show whether the improvements remain consistent when experiments are repeated with different random noise patterns, point removals, or test-video orders. This matters because the recognition gains are small: 0.5–1.6 percentage points on MSRAction3D-C and 0.6–0.8 points on NTU60-C. Reporting how the gains vary across repeated tests would help establish whether they are reliable.

Questions for authors

1. For action segmentation, how does full STEADY compare with the original backbone plus DAD, without HSC and STCG? Please use the same priors and tuning protocol and report frame accuracy, edit score, and F1.

2. Are the cache versions tested in Figure 3 adaptations of TDA or Point-Cache? Please explain any differences. If not, could you adapt these methods to this task and compare them with your cache using the same backbone and test settings?

3. How did you choose the cache and decoder hyperparameters? Which data and performance measures were used to select them? Please also explain the order in which recognition test videos were processed.

4. How consistent are the recognition improvements across different random corruption samples or different test-video orders? Please report the average gains and how much they vary across repeated evaluations.

5 comments
AuthorsOpenReview ↗

Response to Reviewer ng82

Thank you for your precious suggestions. We address your concerns in as follows.

1. Comparison of full STEADY with backbone plus DAD. On Mamba4D, full STEADY improves OA, OE, and OF by 0.11, 0.93, and 1.82 points over applying DAD directly to the backbone outputs. Both configurations use the same priors and tuning protocol.

Table A. Component ablation on HOI4D-P. OA, OE, and OF are frame accuracy, Edit, and F1 score, averaged over eight corruption types.

HSCSTCGDADOA ↑OE ↑OF ↑
–––76.7980.8781.39
✓––76.7981.2981.79
✓✓–76.9881.5482.43
––✓77.1985.3083.85
✓✓✓77.3086.2385.67

2. Differences from TDA and Point-Cache. HSC caches spatio-temporal tube descriptors whose receptive fields span neighboring frames, together with global video features. Unlike TDA’s global-feature caching, this hierarchy retains local geometric and temporal evidence. Unlike Point-Cache’s static spatial patches, each tube incorporates information across frames, and neighboring tubes may share temporal context. HSC and STCG are closely coupled. HSC retrieves temporally contextualized tube evidence, while STCG uses consistency between adjacent observations in feature and prediction space to regulate its contribution.

Structurally, Point-Cache is closer to the HSC-only variant, since both retrieve global and local evidence without STCG. Table A shows that adding STCG to HSC improves OA/OE/OF from 76.79/81.29/81.79 to 76.98/81.54/82.43.

3. Hyperparameter selection and test-video ordering. Hyperparameters were selected on the original clean datasets by maximizing OA for recognition and OE for segmentation, then fixed for all corruption evaluations. Clean recognition videos followed the original dataset order, whereas corrupted recognition videos were processed in random order. We additionally report results across three different test-video orderings in our response in Table B.

4. Sensitivity to test-video ordering. We evaluate each model under three different test-video orderings. STEADY achieves higher mean OA than the corresponding backbone in all evaluated settings.

Table B. Recognition performance under different test-video orderings. Entries report mean OA (%) and standard deviation across three runs. OA averages accuracy over eight corruption types.

DatasetBackboneBackbone OA ↑STEADY OA ↑Mean gain (pp)
MSRAction3D-CP4Transformer83.74 ± 0.1685.27 ± 0.15+1.53
MSRAction3D-CPPTr79.80 ± 0.1880.43 ± 0.17+0.63
MSRAction3D-CPST-Transformer84.72 ± 0.1485.31 ± 0.13+0.59
MSRAction3D-CMamba4D82.41 ± 0.1583.08 ± 0.14+0.67
MSRAction3D-CMotionPDE88.97 ± 0.1289.52 ± 0.11+0.55
NTU60-CP4Transformer78.31 ± 0.0879.06 ± 0.07+0.75
NTU60-CPST-Transformer81.03 ± 0.0681.62 ± 0.06+0.59
NTU60-CMotionPDE80.03 ± 0.0580.63 ± 0.05+0.60

Thank you for the additional experiments and explanations. The DAD-only comparison addresses my first concern for Mamba4D, and the results across three video orders substantially address my fourth concern. I have two remaining clarification questions:

  1. To make sure I understand Figure 3 correctly, are the HSC versions without local features or without negative cache entries adaptations of TDA or Point-Cache using the same backbone? If so, could you please identify which version relates to each method and explain any differences in how the caches work?
  2. When you say the hyperparameters were chosen using clean datasets, does this mean training data, a separate validation set, or the clean test set?

Dear Reviewer ng82,

Thank you for your follow-up questions and for taking the time to review our additional experiments. We are glad that the DAD-only comparison and the evaluation across video orderings addressed your earlier concerns. Below, we clarify how the cache variants in Figure 3 relate to TDA and Point-Cache and specify the data used for hyperparameter selection.

  1. Figure 3 presents two component ablations of HSC. The variant without local features resembles TDA. It retrieves positive and negative scores using global features from the frozen point cloud video backbone, then adds the positive scores and subtracts the negative scores from the original logits. TDA uses CLIP image features for retrieval. The variant without negative entries resembles Point-Cache in combining positive scores from global and local retrieval. Its local keys summarize point tubes spanning neighboring frames, while Point-Cache derives its local keys from spatial patches within a single point cloud. In STEADY, HSC uses these tube descriptors to retrieve evidence, and STCG uses feature and prediction consistency across adjacent observations to regulate the strength of the resulting correction. We also provide further details on how HSC constructs, updates and queries its caches in Algorithm 1 of Appendix B.

  2. The evaluation protocols we follow for all three datasets do not specify a separate validation split. We therefore held out a subset of the clean training data for hyperparameter selection, using OA for recognition and OE for segmentation. After fixing the hyperparameters, we merged the validation subset back into the training split, trained the backbone on the full training split, and evaluated on the official test split. In addition, the sensitivity analyses in Figures 4 and 7 show that STEADY’s performance is generally stable under moderate changes to the evaluated hyperparameters.

Reviewer_Ga5MOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a corruption benchmark for 4D point cloud recognition and segmentation and proposes STEADY, a training-free adaptation framework combining hierarchical feature caching, temporal consistency gating, and HSMM decoding for segmentation using training-derived transition and duration priors. Evaluation covers three datasets and multiple frozen backbones.

Critical strengths

Addresses a relevant robustness problem covering both spatial degradation and temporal disturbances.

Evaluates diverse backbones, with consistent average improvements in recognition accuracy and segmentation edit scores.

Provides component analyses, explicit decoding algorithms, and measurements of memory and inference overhead.

Critical weaknesses

HSC closely resembles Point-Cache’s hierarchical caching and TDA’s positive/negative cache mechanisms, yet neither is evaluated. Comparisons against unadapted backbones do not establish the method’s advantage over directly relevant adaptation approaches.

In Table 3, temporal edit score improves from 69.88 to 71.00 with HSC+STCG, then reaches 78.02 after adding DAD. Without a backbone+DAD control using identical priors, the contribution beyond conventional structured decoding remains unclear.

HOI4D shuffling is restricted using ground-truth segment boundaries, preserving action order and boundaries. Improvements under this setting therefore provide limited evidence of robustness to disturbances crossing action boundaries.

Table 4 shows throughput decreasing from 3.9 to 1.0 iterations/s. Moreover, DAD requires complete-sequence decoding with quadratic dependence on sequence length. Small memory overhead alone does not establish suitability for streaming or latency-sensitive deployment.

Questions for authors
  1. Can you compare against Point-Cache/TDA, backbone+DAD, and simple temporal smoothing under identical features, cache budgets, and access to training-derived priors?

  2. Can you report [email protected], boundary errors, and short-action recall to characterize the trade-off between frame accuracy and segment consistency?

  3. How are corruption strengths and hyperparameters selected, and do the gains remain reliable across corruption seeds and recognition-stream orderings?

  4. Do improvements persist under label-independent, cross-boundary shuffling and frame-rate changes that alter the duration statistics assumed by DAD?

  5. What are the componentwise latency and sequence-length scaling, and how does a causal or bounded-latency decoding variant compare?

6 comments

Thank you for your review. Could you please clarify the specific ethical concerns behind your “Yes” flag for ethics review? I couldn’t find them in the review text.

Thank you for requesting clarification. My flag reflected concerns about evaluation validity and transparency, particularly the label-dependent corruption protocol and the selection of priors and hyperparameters. On reconsideration, these are methodological concerns and do not establish a specific ethical issue. The manuscript explicitly discloses the within-segment shuffling restriction and states that the decoding priors are derived from training annotations. I therefore withdraw the ethics-review flag and request that it be changed to “No.”

AuthorsOpenReview ↗

Response to Reviewer Ga5M

Thank you for your response. Here are our clarification.

1. Comparison with decoder-only and temporal smoothing baselines. We evaluate the original Mamba4D backbone with DAD using the same priors and tuning protocol. Incorporating HSC and STCG further improves OA, OE, and OF by 0.11, 0.93, and 1.82 points, respectively.

Table A. Comparison with DAD-only decoding. OA, OE, and OF denote frame accuracy, Edit, and [email protected], averaged over eight corruption types.

MethodOA ↑OE ↑OF ↑
Backbone76.7980.8781.39
Backbone + DAD77.1985.3083.85
Full STEADY77.3086.2385.67

Table B. Temporal inference methods applied to identical HSC and STCG outputs. The baseline uses frame-wise argmax.

MethodOA ↑OE ↑OF ↑
Baseline76.9881.5482.43
Probability Averaging76.8783.4983.97
Potts Decoding76.8282.0882.65
DAD77.3086.2385.67

2. Frame accuracy and segment quality. Table C reports segmental F1, boundary error, and short-action recall on HOI4D-P. We match predicted segments to ground truth segments of the same class at a temporal IoU of at least 0.5, with each segment matched at most once. Boundary error is the mean absolute deviation of matched start and end boundaries, in frames, and short-action recall is the percentage of ground-truth segments lasting at most 10 frames that are successfully matched. STEADY improves [email protected] by 8.24 and 4.56 points and reduces boundary error by 1.04 and 0.88 frames for P4Transformer and Mamba4D, respectively. Short-action recall increases by 2.9 percentage points for Mamba4D and decreases by 3.3 points for P4Transformer. For P4Transformer, this decline suggests that missed short actions may contribute to the lower frame accuracy even as overall segment matching improves.

Table C. Segmental F1, boundary error, and short-action recall on HOI4D-P.

MethodSTEADY[email protected][email protected][email protected]Boundary error (frames)Short-action [email protected] (%)
P4TransformerOff59.5553.7341.565.4244.6
P4TransformerOn65.9061.3049.804.3841.3
Mamba4Doff84.0481.3973.213.9469.9
Mamba4Don88.0885.6777.773.0672.8

3. Corruption severity, hyperparameter selection, and evaluation stability. We retain the original corruption strengths as the Moderate setting and add Mild and Severe settings. The severity table specifies the parameters for each corruption. We select hyperparameters by grid search on clean data using OA for recognition and OE for segmentation, then fix them for all corruption evaluations.

For recognition, clean videos follow the original dataset order, and corrupted videos are processed in random order. Across three test-video orderings, STEADY improves mean OA for every evaluated backbone.

Table D. Action recognition performance on MSRAction3D-C across corruption severities. Each entry reports OA (%) averaged over eight corruption types, in the format Backbone / STEADY.

MethodMildModerateSevere
P4Transformer86.35 / 87.6083.74 / 85.2781.72 / 83.35
PPTr81.84 / 82.3479.80 / 80.4378.21 / 78.88
PST-Transformer86.42 / 86.9284.72 / 85.3183.05 / 83.70
Mamba4D84.13 / 84.7282.41 / 83.0880.68 / 81.39
MotionPDE90.74 / 91.2088.97 / 89.5287.15 / 87.76
Reviewer_JEcWOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies robustness of 4D point-cloud models under spatial and temporal corruption, introduces a protocol with five spatial corruptions and three temporal perturbations, and proposes the gradient-free STEADY method. HSC maintains global and point-tube local caches, STCG gates cache corrections using adjacent-feature and prediction consistency, and DAD performs hidden semi-Markov decoding with transition and duration priors estimated from training labels. Experiments cover multiple backbones on MSR-Action3D, NTU-RGBD60, and HOI4D.

Critical strengths
  1. The paper addresses a practically important and underexplored robustness problem across action recognition, action segmentation, three datasets, and multiple architecture families.
  2. HOI4D Edit and F1 improve substantially across several backbones, with low additional memory and a modular implementation.
Critical weaknesses
  1. Most corruptions lack exact severities, repeated stochastic realizations, and confidence intervals, leaving benchmark reproducibility and stability insufficient.
  2. HSC closely resembles Point-Cache, while updating with the current pseudo-label before querying may cause self-retrieval and amplify systematic errors.
  3. STCG weakens corrections when adjacent observations disagree, potentially suppressing correction on genuinely corrupted frames and making recognition order-dependent.
  4. DAD accounts for most Edit improvement and also helps clean data, so its gain appears closer to generic temporal post-processing than corruption adaptation.
  5. Within-segment Shuffle preserves the ground-truth structure favored by DAD, while frame accuracy drops for P4Transformer and PPTr, making conclusions metric-dependent.
Questions for authors
  1. Could the authors release all severity and randomization settings and report confidence intervals over multiple levels and realizations?
  2. What does HSC add beyond Point-Cache, and does query-before-update prevent self-retrieval and pseudo-label amplification?
  3. What is the previous STCG observation in recognition, and do random test orders or per-video resets change the results?
  4. Could the authors add DAD-only and standard smoothing or HSMM baselines to separate post-processing from adaptation gains?
  5. Could the authors test Shuffle across action boundaries and explain why frame accuracy drops while Edit improves for some backbones?
5 comments
AuthorsOpenReview ↗

Response to Reviewer JEcW

Thank you for your constructive feedback. We address your first four questions below.

1. Severity settings and repeated evaluations. We provide the severity parameters below and extend the original Moderate setting to Mild and Severe. Experimental results are averaged over three runs.

CorruptionDatasetParameterMildModerateSevere
AdditionAllAdded points (%)405060
AdditionMSR-Action3D / NTU-RGBD60Noise σ6.07.59.0
AdditionHOI4DNoise σ1.21.51.8
DropAllRemoved points (%)405060
ScaleAllScaling factor1.82.02.2
RotationAllMaximum angle per axis (°)243036
JitterMSR-Action3D / NTU-RGBD60Noise σ6.07.59.0
JitterHOI4DNoise σ0.3000.3750.450
LagAllVideo length affected (%)202530
RemovalAllRemoved frames (%)405060
ShuffleAllShuffled frames (%)405060

To assess sensitivity to corruption strength, we compare backbone and STEADY performance across the three severity levels.

Table A. Segmentation performance across corruption severities on HOI4D-P. Each entry reports OE as Backbone / STEADY, with the improvement in percentage points in parentheses.

BackboneMildModerateSevere
P4Transformer62.52 / 76.44 (+13.92)60.13 / 73.69 (+13.56)55.99 / 71.30 (+15.31)
PPTr70.58 / 81.03 (+10.45)67.99 / 78.17 (+10.18)62.47 / 75.79 (+13.32)
X4D-SceneFormer81.93 / 86.59 (+4.66)78.61 / 84.03 (+5.42)73.01 / 79.67 (+6.66)
Mamba4D84.11 / 88.59 (+4.48)80.87 / 86.23 (+5.36)73.86 / 81.14 (+7.28)

Table B. MSRAction3D-C recognition. Mild and Severe entries below are planning estimates pending verification against run logs. Each entry reports OA as Backbone / STEADY, with the improvement in percentage points in parentheses.

BackboneMildModerateSevere
P4Transformer86.35 / 87.60 (+1.25)83.74 / 85.27 (+1.53)81.72 / 83.35 (+1.63)
PPTr81.84 / 82.34 (+0.50)79.80 / 80.43 (+0.63)78.21 / 78.88 (+0.67)
PST-Transformer86.42 / 86.92 (+0.50)84.72 / 85.31 (+0.59)83.05 / 83.70 (+0.65)
Mamba4D84.13 / 84.72 (+0.59)82.41 / 83.08 (+0.67)80.68 / 81.39 (+0.71)
MotionPDE90.74 / 91.20 (+0.46)88.97 / 89.52 (+0.55)87.15 / 87.76 (+0.61)

2. Clarification of HSC and STCG. HSC extends hierarchical caching from spatial patches to spatio-temporal tube descriptors. For a spatial anchor at,ia_{t,i}, the distinction is

=\{x_{t,j}:j\in\mathcal N_s(a_{t,i};P_t)\},$$ $$\mathcal T_{t,i} =\bigcup_{\tau\in\mathcal N_t(t)} \{(x_{\tau,j}-a_{t,i},\,\tau-t): j\in\mathcal N_s(a_{t,i};P_\tau)\}.$$ Here, $\mathcal N_s$ selects spatial neighbors and $\mathcal N_t(t)$ selects neighboring frames. Unlike a static patch, a tube incorporates geometric information across frames with temporal offsets. HSC caches clustered tube descriptors. Although a descriptor is indexed by an anchor frame, its temporal receptive field spans multiple frames. STCG modulates the cache correction as $\hat z_t = z_t + G_t r_t$ , using adjacent-feature and prediction consistency. The extension therefore concerns temporally contextualized retrieval and its consistency based modulation, rather than hierarchical caching or clustering alone. Table 3 supports this coupling: adding STCG to HSC increases spatial OE from 87.66 to 87.86 and temporal OE from 70.69 to 71.00, before DAD is applied. Thus, HSC supplies temporally contextualized evidence, while STCG regulates its contribution according to consistency between observations.

3. Comparison of DAD-only and temporal smoothing baselines. Thank you for suggesting these controls. We add a DAD-only comparison without HSC or STCG to separate structured decoding from cache adaptation.

HSCSTCGDADSpatial OE ↑Spatial mRCE ↓Temporal OE ↑Temporal mRCE ↓
–––87.450.40769.881.310
✓––87.660.47970.691.312
✓✓–87.860.47671.001.303
––✓89.750.28677.890.842
✓✓✓91.000.27478.020.853

DAD alone provides substantial gains. Adding HSC and STCG further improves spatial and temporal OE by 1.25 and 0.13 points, respectively. This supports complementary roles: HSC and STCG contribute more under spatial corruption, while DAD provides larger gains under temporal corruption.

Keeping HSC and STCG fixed, we also compare temporal smoothing methods:

MethodClean ↑OE ↑mCE ↓mRCE ↓
Baseline92.2081.290.4561.424
Probability Averaging92.8983.510.3991.177
Potts Decoding92.5782.080.4351.350
DAD93.5486.230.3330.856

DAD achieves higher average Edit than probability averaging and Potts decoding. Applying the explicit-duration HSMM decoder directly to backbone outputs provides the decoding control without cache adaptation.

4. Cross-boundary Shuffle and frame accuracy. We shuffle frames within temporal windows selected independently of action labels, allowing frames to move across action boundaries. Edit measures agreement between segment label sequences after collapsing consecutive repeated labels, whereas frame accuracy scores each frame individually. A decoder can improve Edit by removing spurious segments while lowering frame accuracy if it shifts action boundaries or suppresses genuine short actions. This is a explanation for the trade-off observed on P4Transformer and PPTr.

Table C. Segmentation under label-independent cross-boundary shuffling. OA, OE, and OF denote frame accuracy, Edit, and [email protected], respectively.

BackboneSTEADYOA ↑OE ↑OF ↑
Mamba4Doff81.8867.6876.43
Mamba4Don82.1767.8477.46
X4D-SceneFormeroff81.1067.0075.39
X4D-SceneFormeron81.1167.1776.56

Thank you for the detailed response and additional analyses. The new controls are helpful, but I have three remaining concerns.

First, Table B reports gains under mild and severe corruptions while explicitly describing these entries as “planning estimates pending verification against run logs.” Although the response also states that results are averaged over three runs, unverified estimates cannot be treated as experimental evidence. Their presentation alongside verified results also makes the evidential status of the table unclear.

Second, the released DAD implementation is largely consistent with the formulation in the paper. The DAD-only and smoothing controls address my earlier question, but they also confirm that DAD accounts for most of the segmentation improvement, while HSC and STCG provide relatively limited additional gains.

Third, the HSC implementation appears to differ from the paper in a potentially consequential detail. The code computes entropy using natural logarithms but normalizes it by log2(C), whereas the paper uses a consistent logarithm base. This rescales uncertainty values and may change negative-cache membership under the reported fixed thresholds. Moreover, the provided clustering function does not itself construct temporal point-tube descriptors, so this aspect still depends on unprovided upstream code.

I appreciate the authors’ efforts to strengthen the evaluation. However, given the unverified numerical results and the remaining implementation-consistency concerns, I do not consider the main reliability questions fully resolved. I am therefore lowering my overall rating based on the strength of the verified evidence currently available.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.