acceptodds
Under review as a conference paper at ICLR 2027

Active Test-time Adaptation for Temporal Action Segmentation

Abstract

Temporal action segmentation (TAS) models often suffer from performance degradation during deployment due to temporal distribution shifts and boundary ambiguity. Active test-time adaptation has shown promising results in image recognition by strategically querying sparse human feedback for adaptation. However, extending this paradigm to TAS is challenging due to long temporal span and action boundary ambiguity in procedural videos. To address these challenges, we present the first active test-time adaptation framework for TAS to strategically query sparse human annotations to ground the adaptation process. Our approach is based on a principled analysis of prediction dynamics, revealing that model hesitation under temporal smoothing constraints manifests as informative entropy peaks at action boundaries. We then propose directional boundary overwrite with agnostic and semantic variants that convert local feedback into adaptation labels, offering different trade-offs between annotation effort and supervision reliability. Our results demonstrate that annotating a few key temporal anchors significantly improves performance, establishing a new foundation for active adaptation of TAS models.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
video understanding, temporal action segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_voq1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

per proposes an active test-time adaptation framework for temporal action segmentation (TAS). The method selects informative predicted action boundaries for sparse annotation and introduces agnostic and semantic directional overwrite strategies to construct adaptation labels. The model is subsequently adapted using sparse classification supervision and temporal smoothing. Experiments on Breakfast, GTEA, and 50Salads with MS-TCN and ASFormer demonstrate overall performance improvements under limited annotation budgets in the evaluated transductive setting.

Critical strengths

1.The paper addresses a meaningful problem by introducing active human feedback into test-time adaptation for temporal action segmentation. The boundary-centric query strategy is motivated by the temporal structure of TAS, and the directional overwrite mechanisms provide a practical way to construct supervision from sparse feedback.

2.The experimental evaluation covers three established TAS benchmarks and two representative backbones. Comparisons with active selection and adaptation baselines, annotation-budget analyses, and a user study provide useful empirical evidence for the effectiveness and practical potential of sparse boundary supervision in the evaluated setting.

Critical weaknesses

1.Potentially incorrect supervision induced by directional boundary overwrite. Equations (11) and (13) assign the same adaptation label to both frames adjacent to every queried predicted boundary. However, if a predicted boundary already coincides with the ground-truth transition, these two frames should have different labels under the standard frame-wise segmentation protocol. Consequently, even with correct oracle feedback, the overwrite rule necessarily assigns an incorrect target to at least one of the two frames. This raises a fundamental concern about the reliability of the proposed adaptation supervision. The paper does not quantify how often this situation occurs or whether the benefits of local label consistency outweigh the resulting supervision errors.

2.Insufficient validation of the distribution-shift motivation. The paper emphasizes temporal distribution shifts during deployment, but the main experiments evaluate adaptation on standard test splits without explicitly examining controlled temporal shifts or cross-domain scenarios. The results support the effectiveness of sparse test-set supervision in the evaluated setting, but do not sufficiently establish robustness under the distribution changes motivating the broader deployment claims. The transductive adaptation protocol itself is valid; the concern is the limited empirical support for the broader deployment claims.

Questions for authors

1.When a predicted boundary coincides with the ground-truth transition, Eqs. (11) and (13) still assign the same label to both adjacent frames, making at least one adaptation target incorrect under frame-wise ground-truth labels. How does the proposed method address this issue? Could the authors quantify the correctness of the generated adaptation labels for both agnostic and semantic overwrite, and clarify whether preserving correctly localized boundaries would affect the reported results?

2.How robust is the proposed framework to incorrect human feedback? Given that the main experiments use simulated oracle responses and the user study evaluates annotation time rather than feedback accuracy, can the authors report results using actual human annotations or analyze the effect of imperfect feedback?

3.Can the authors provide experimental evidence under explicit temporal distribution shifts, such as changes in action-duration distributions or cross-domain evaluation, to substantiate the claimed deployment robustness?

1 comment

W1 & Q1. Already-correct boundaries

Thank you for asking about this detail. The short answer is: our implementation imposes no agnostic overwrite when a predicted boundary matches the ground truth. We did not mention this in the paper for two reasons: 1) the case is extremely rare (<1%), and 2) judging the exact end frame of an action is ill-posed for annotators.

Specifically, over 99% of predicted boundaries across all datasets do not coincide exactly with the ground-truth transition, so nearly every query concerns a boundary that does need to move. The exact-match case is very rare, which is why we did not explicitly discuss it in the submission and focused instead on the far more common case of misaligned boundaries. In our implementation, when a predicted boundary does coincide exactly with the ground truth, it is kept as it is and no agnostic overwrite is applied.

Moreover, the directional query is closer to how annotation works in practice. Action transitions are ambiguous at the frame level, so asking an annotator whether a predicted boundary is exactly right is an ill-posed question, whereas judging whether it comes too early or too late is easy and consistent.

We will mention how we handle it explicitly in the revision.


W2 & Q3. Validation of the shift motivation

Unfortunately, we are unable to evaluate under a cross-domain shift: TAS models are trained on a fixed set of action classes, and we could not find videos recorded under a different distribution that share the action classes of these benchmarks.

Although training and test videos come from the same benchmark, the standard splits already contain a substantial temporal shift. We measured it on Breakfast:

  • Action duration: 37.2% of action classes differ in mean duration between training and test by a factor of at least 1.25, and 15.8% by a factor of at least 1.5.
  • Temporal pattern: 42.1% of test videos follow an action ordering that never occurs in the training set.

The test sets therefore substantially differ from the training sets in both action durations and temporal patterns, and the improvement validate the effectiveness of our approach.


Q2. Noisy feedback

Thank you for your great question. We would like to first point out that simulating the annotator with ground-truth labels follows the standard protocol in existing active test-time adaptation (what SimATTA, BiTTA also adopt). A full human-in-the-loop evaluation is prohibitive at this scale: it would need human answers at every queried boundary across all test videos, splits, backbones and settings.

We therefore simulated imperfect overwrite on Breakfast with MS-TCN and agnostic overwrite. For each video, we randomly flip 10%, 20%, 30% of the overwrite direction of queried boundaries and report the performance in the table below. As we can see, performance generally declines as the error rate rises, but the method still manages to stay above the pretrained model. The method is therefore robust to moderate annotation error.

SettingAccEditF1@10F1@25F1@50
Pretrained65.467.568.563.049.6
Ours, no errors73.773.274.369.957.6
Ours, 10% errors72.572.373.668.856.0
Ours, 20% errors72.471.872.668.155.4
Ours, 30% errors71.270.271.466.552.5
Reviewer_A5bZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies active test-time adaptation for temporal action segmentation. The authors propose selecting predicted action boundaries for sparse human annotation and introduce two directional boundary overwrite strategies. Agnostic overwrite uses boundary feedback to adjust predicted labels, while semantic overwrite additionally corrects action categories. The resulting supervision is combined with temporal smoothing to adapt the segmentation model. Experiments on Breakfast, GTEA, and 50Salads show improvements over several active sampling and test-time adaptation baselines.

Critical strengths
  1. The paper considers a practical problem: adapting temporal segmentation models with limited human feedback.
  2. Boundary‑based querying is a reasonable way to reduce annotation requirements for long videos.
  3. The distinction between agnostic and semantic overwrite provides a useful trade‑off between annotation effort and supervision quality.
  4. Experiments cover three datasets and two common TAS architectures, MS‑TCN and ASFormer.
  5. The annotation‑time study and budget analysis help assess the practical cost of the proposed approach.
Critical weaknesses
  1. The methodological novelty is limited. The work mainly extends active test‑time adaptation methods such as SimATTA and BiTTA to temporal segmentation. Compared with active TAS (Su and Elhamifar, 2024), the main distinction is the test‑time setting and boundary overwrite strategy.
  2. The distribution‑shift motivation is not adequately tested. Experiments use standard TAS benchmarks without explicitly evaluating changes in recording conditions, action durations, or temporal patterns.
  3. The analysis relies on simplified assumptions. The connection between temporal smoothing and boundary entropy is derived under a two‑class transition assumption. Its applicability to more ambiguous multi‑class boundaries remains unclear.
  4. The evaluation uses simulated annotation feedback. The user study measures annotation time, but does not establish how reliably humans provide boundary corrections. The reported adaptation gains may therefore be optimistic.
  5. The selection strategy cannot directly recover missed boundaries. Since candidate queries come from predicted boundaries, the method may struggle when the initial model merges adjacent actions or misses an action entirely.
  6. The contribution of the proposed overwrite rule is not fully isolated. It is unclear how much of the improvement comes from boundary selection, label correction, or ordinary supervised adaptation using sparse test‑time labels.
Questions for authors
  1. The paper motivates active adaptation through temporal distribution shifts, but evaluates primarily on standard TAS splits. Could the authors provide evidence that the method remains effective under explicit shifts in action duration, visual appearance, or temporal structure?
  2. The adaptation experiments simulate human feedback using ground‑truth annotations. How sensitive are the reported improvements to incorrect or imprecise boundary judgments? Could the authors evaluate adaptation with noisy annotations or actual user‑provided corrections?
  3. Since query candidates are restricted to predicted action boundaries, how does the method handle missing boundaries or incorrectly merged action segments? Could the authors report performance separately for initially correct, shifted, and missing boundaries?
  4. How much of the improvement comes specifically from directional boundary overwrite? A comparison with ordinary sparse supervised adaptation using the same queried frames and annotation budget would help establish its contribution.
7 comments

W1. Limited novelty over SimATTA, BiTTA and active TAS

Our work indeed study the same setup (active TTA) as SimATTA and BiTTA. However, the challenges it addresses and methods we propose are specific to TAS and have not been studied in those works.

SimATTA and BiTTA query individual samples and ask about a class label or if the label is correct. Our method queries predicted boundaries and asks a directional question about the transition, which has no counterpart in image-level active TTA. The agnostic variant needs no class label at all, and it converts a single answer into supervision through the boundary overwrite. We further propose two overwrite variants to account for the annotation efforts which have not been studied in existing SimATTA nor BiTTA.

With respect to active TAS, their selection relies on action prototypes learned with contrastive objectives during training; it might be challenging for a deployed model trained without such prototypes (Appendix E). Our selection needs only the entropy of the model’s own predictions.

To reiterate, our contributions are 1) an efficient active frame query selection mechanism from the model effect from the smoothing loss in training, 2) two variants of overwrite mechanism that provides trade-offs between annotation cost and adaptation performance, 3) an effective adaptation strategy that leverages the sparse supervision. These contributions are distinct from the above methods and should not be considered as an ‘extension’ of them.


W2 & Q1. Distribution shift, evidence under shifts in duration, appearance or temporal structure.

Although training and test videos come from the same benchmark, the standard splits already contain a substantial temporal shift. We measured it on Breakfast:

  • Action duration: 37.2% of action classes differ in mean duration between training and test by a factor of at least 1.25, and 15.8% by a factor of at least 1.5.
  • Temporal pattern: 42.1% of test videos follow an action ordering that never occurs in the training set.

The test sets therefore substantially differ from the training sets in both action durations and temporal patterns, and the improvement validate the effectiveness of our approach.


W3. Two-class assumption

Thank you for your comment. As we mentioned in Appendix A: Our analysis of smoothing loss is intended as intuition that motivates our boundary selection in general, not as a formal proof of optimality.

Concretely, the two-class case is used because it matches how ground truth is defined in TAS: each boundary marks transition from one action class to another (L. 215). And this aligns closely with what we observe (L.634-636): on Breakfast, over 90% of frames have a top-1 + top-2 probability above 0.9 (96% above 0.8).

Furthermore, the 3rd predicted boundary in Fig. 4 shows semantic confusion between sequeeze_orange, cut_orange, and take_glass, yet the local peak is still around the label transition point.


W4 & Q2. Noisy feedback

We would like to first point out that simulating the annotator with ground-truth labels follows the standard protocol in existing active test-time adaptation (what SimATTA, BiTTA also adopt). A full human-in-the-loop evaluation is prohibitive at this scale: it would need human answers at every queried boundary across all test videos, splits, backbones and settings.

We therefore simulated imperfect overwrite on Breakfast with MS-TCN and agnostic overwrite. For each video, we randomly flip 10%, 20%, 30% of the overwrite direction of queried boundaries and report the performance in the table below. As we can see, performance generally declines as the error rate rises, but the method still manages to stay above the pretrained model. The method is therefore robust to moderate annotation error.

SettingAccEditF1@10F1@25F1@50
Pretrained65.467.568.563.049.6
Ours, no errors73.773.274.369.957.6
Ours, 10% errors72.572.373.668.856.0
Ours, 20% errors72.471.872.668.155.4
Ours, 30% errors71.270.271.466.552.5

W5 & Q3. Missing boundaries can not be directly recovered.

To measure this, we align the ordered segment before and after adaptation with respect to the ground truth segment sequence. We report two criteria in the tables below: a relaxed one, where a match on at least one side counts, and a strict one, where the full transition must match, each at several distance tolerances.

As we can see, under the relaxed criterion, adaptation raises the share of matched boundaries by 4.3 points, from 88.9% to 93.2% (7.4% recovered, 3.1% lost). Under the strict criterion, it raises it by 7.6 points, from 49.6% to 57.2% (17.9% recovered, 10.3% lost). The gain holds at every tolerance, for example from 52.1% to 58.3% (relaxed) and from 37.1% to 43.4% (strict) within 50 frames, and in every setting more boundaries are recovered than lost.

Relaxed: at least one side aligned counts as positive

ToleranceBefore %Recovered%Lost %After %
Any distance88.97.43.193.2
≤10 frames19.515.112.622.1
≤20 frames33.118.215.136.2
≤50 frames52.117.611.458.3

Strict: full transition required

ToleranceBefore %Recovered %Lost %After %
Any distance49.617.910.357.2
≤10 frames15.212.310.017.5
≤20 frames25.015.411.928.5
≤50 frames37.116.610.343.4

W6 & Q4. Isolating the overwrite

The contribution of overwrite is validated in the table below. With no overwrite (uses predicted labels), the performance of sparse supervised adaptation actually falls below the unadapted model in Acc (62.7 vs. 65.4). It is likely because training on these unreliable/ambiguous preidcited labels can potentially harm the discriminability of the classifier. While adding the directional answer on the same frames raises accuracy by 11.0 points, Edit by 5.4 and F1@50 by 10.3. The improvement is therefore attributable to the directional feedback and not to sparse supervision at boundary frames in itself.

SettingAccEditF1@10F1@25F1@50
pretrained65.467.568.563.049.6
No overwrite62.767.868.662.147.3
Ours73.773.274.369.957.6

Dear reviewer A5bZ,

As the discussion period is ending soon, we are writing to ask whether you have had a chance to look at our rebuttal.

We would be glad to answer any remaining questions before the discussion closes. If our rebuttal has addressed your concerns, we would appreciate it if you update your assessment.

Thanks!

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

This paper proposes an active test-time adaptation framework for Temporal Action Segmentation (TAS) using sparse human annotations. The method selects informative action boundaries based on prediction entropy and introduces agnostic and semantic directional boundary overwrite to generate adaptation supervision. Experiments on Breakfast, GTEA, and 50Salads demonstrate promising improvements with minimal annotation budgets.

Critical strengths

Novel problem setting: Extending active test-time adaptation to TAS is interesting and practically relevant. Efficient supervision: Boundary-centric querying and directional overwrite offer an intuitive approach to reducing annotation costs. Promising results: Consistent improvements across multiple benchmarks and two backbones demonstrate the potential of sparse boundary supervision.

Critical weaknesses

Limited theoretical justification: The entropy analysis relies on simplified temporal smoothing assumptions and does not rigorously explain why medium-entropy boundaries are optimal. Potentially incorrect supervision: Directional overwrite assigns identical labels to adjacent frames, potentially introducing errors when predicted boundaries already match ground truth. Insufficient ablation studies: The individual contributions of boundary selection, directional overwrite, and temporal smoothing are not adequately isolated. Limited real-world validation: Experiments rely on simulated ground-truth feedback without evaluating robustness to annotation errors or substantial temporal distribution shifts.

Dependence on initial predictions: Missing or severely mislocalized boundaries cannot be effectively corrected, potentially limiting adaptation under challenging conditions.

Questions for authors

How does directional overwrite handle already-correct boundaries? Would introducing a no-overwrite option prevent incorrect supervision? Can the authors demonstrate a quantitative relationship between boundary entropy, localization error, and adaptation effectiveness to justify medium-entropy selection? Can the authors provide controlled ablations isolating boundary selection, directional overwrite, and temporal smoothing under identical supervision budgets? How does the framework perform when initial predictions miss entire action segments or contain substantial boundary errors? Can the authors evaluate adaptation with noisy human feedback and controlled temporal distribution shifts?

2 comments

W1 & Q2. Smoothing loss analysis does not rigorously justify medium-entropy selection.

The smoothing loss analysis is not intended to justify the medium-entropy selection. Our analysis of smoothing loss in Sec 3.2 is intended as intuition that motivates our boundary selection strategy. What the analysis suggests is that predicted boundaries can be used as the candidate set for active query (see L.196).

The choice of medium-entropy boundaries within this set is made separately in Sec. 3.3. They are selected by their relative entropy rankings among the boundaries (see L. 237-256). The ranking only compares entropies across boundaries and does not rely on the smoothing-loss analysis.

To support the choice of medium entropy empirically, we measure localization and transition correctness per entropy group on Breakfast with MS-TCN (table below). For each group, we report the share of predicted boundaries that fall within 10, 20, and 50 frames of the nearest ground-truth boundary, and the share whose predicted transition is correct. Both measures decrease monotonically as entropy increases.

Entropy< 10 frames< 20 frames< 50 framesTransition correct
High17.531.653.733.2
Medium18.132.054.739.3
Low22.437.759.749.7

Low-entropy boundaries are the best localized (22.4% within 10 frames) and have the most reliable transition labels (49.7%), so they leave the least room for correction. High-entropy boundaries leave the most room (17.5% within 10 frames) but have the least reliable labels (33.2%), which makes them risky to adapt on. Medium-entropy boundaries offer the best tradeoff: they leave almost as much room for boundary adjustment as high-entropy ones (18.1% vs. 17.5%), with noticeably more reliable labels (39.3% vs. 33.2%).

The effectiveness of the selection is already presented in Table 4, which shows all ranges would help improve the performance but the medium-entropy has the highest boost.


W2 & Q1. Overwrite with already-correct boundaries

Thank you for asking about this detail. The short answer is: our implementation imposes no agnostic overwrite when a predicted boundary matches the ground truth. We did not mention this in the paper for two reasons: 1) the case is extremely rare (<1%), and 2) judging the exact end frame of an action is ill-posed for annotators.

Specifically, over 99% of predicted boundaries across all datasets do not coincide exactly with the ground-truth transition, so nearly every query concerns a boundary that does need to move. The exact-match case is very rare, which is why we did not explicitly discuss it in the submission and focused instead on the far more common case of misaligned boundaries. In our implementation, when a predicted boundary does coincide exactly with the ground truth, it is kept as it is and no agnostic overwrite is applied.

Moreover, the directional query is closer to how annotation works in practice. Action transitions are ambiguous at the frame level, so asking an annotator whether a predicted boundary is exactly right is an ill-posed question, whereas judging whether it comes too early or too late is easy and consistent.

We will mention how we handle it explicitly in the revision.


W3 & Q3. Component ablations on boundary, overwrite and smoothing

Thank you for your comment. The contribution of selection is already in the main paper (Tables 1 and 4) where we compare with other frame selection strategies.

We next provide two new studies for directional overwrite and temporal smoothing. All results are on Breakfast with MS-TCN and agnostic overwrite, at the same query budget.

The contribution of overwrite is validated in the table below. With no overwrite (uses predicted labels), the performance falls below the unadapted model in Acc (62.7 vs. 65.4). It is likely because training on these unreliable/ambiguous labels would potentially harm the discriminability of the classifier. Adding the directional answer on the same frames raises accuracy by 11.0 points, Edit by 5.4 and F1@50 by 10.3. The improvement is therefore attributable to the directional feedback and not to sparse supervision at boundary frames in itself.

SettingAccEditF1@10F1@25F1@50
pretrained65.467.568.563.049.6
No overwrite62.767.868.662.147.3
Ours73.773.274.369.957.6

We further study the contribution of smoothing, as we can see adding the smoothing loss during adaptation boosts every metric by a margin: 0.7 in Acc, 1.2 in Edit and up to 1.8 in F1. This improvement shows that smoothing loss is a useful regulariser.

SettingAccEditF1@10F1@25F1@50
w/o Lsm73.072.073.668.156.3
Ours73.773.274.369.957.6

W4 & Q5. Noisy feedback and temporal shift

Noisy feedback. We would like to first point out that simulating the annotator with ground-truth labels follows the standard protocol in existing active test-time adaptation (what SimATTA, BiTTA also adopt). A full human-in-the-loop evaluation is prohibitive at this scale: it would need human answers at every queried boundary across all test videos, splits, backbones and settings.

We therefore simulated imperfect overwrite on Breakfast with MS-TCN and agnostic overwrite. For each video, we randomly flip 10%, 20%, 30% of the overwrite direction of queried boundaries and report the performance in the table below. As we can see, performance generally declines as the error rate rises, but the method still manages to stay above the pretrained model. The method is therefore robust to moderate annotation error.

SettingAccEditF1@10F1@25F1@50
Pretrained65.467.568.563.049.6
Ours, no errors73.773.274.369.957.6
Ours, 10% errors72.572.373.668.856.0
Ours, 20% errors72.471.872.668.155.4
Ours, 30% errors71.270.271.466.552.5

Temporal shift. Unfortunately, we are unable to evaluate under a cross-domain shift: TAS models are trained on a fixed set of action classes, and we could not find videos recorded under a different distribution that share the action classes of these benchmarks.

We'd also like to add that the standard splits from the benchmark already contain a substantial temporal shift that our work aims to address. For example, on Breakfast:

  • Action duration: 37.2% of action classes differ in mean duration between training and test by a factor of at least 1.25, and 15.8% by a factor of at least 1.5.
  • Temporal pattern: 42.1% of test videos follow an action ordering that never occurs in the training set.

W5 & Q4. Harder cases, missed or mislocalized boundaries

Harder cases. We acknowledge that there can be harder cases and we see this as a trade-off between performance and annotation budget, as discussed in our analysis in Sec. 4.3 (L.415-450).

Our current experiments operate under an extremely limited annotation budget, see how the model works with minimal annotation cost. However, we do have two designs to account for the harder cases. 1) an adjustable boundary selection ratio η\eta and 2) the semantic overwrite variant that provides additional action class labels. Table 3 already shows that allowing larger budget with semantic overwrite can significantly improve the performance.

We further report experiment on Breakfast with asformer below. Increasing the labelling to 0.3%, it further increase the performance, with Acc close to full, showing potential of resolving harder cases with more annotations.

BreakfastFrames labelledAccEditF1@10F1@25F1@50
Pretrained–72.875.276.271.058.9
Ours (sem)0.15%79.177.480.276.063.0
Ours (sem)0.3%83.478.480.978.168.6
Full100%84.582.384.881.672.6

Performance under missed or far-off boundaries. To measure this, we align the ordered segment before and after adaptation with respect to the ground truth segment sequence. We report two criteria in the tables below: a relaxed one, where a match on at least one side counts, and a strict one, where the full transition must match, each at several distance tolerances.

As we can see, under the relaxed criterion, adaptation raises the share of matched boundaries by 4.3 points, from 88.9% to 93.2% (7.4% recovered, 3.1% lost). Under the strict criterion, it raises it by 7.6 points, from 49.6% to 57.2% (17.9% recovered, 10.3% lost). The gain holds at every tolerance, for example from 52.1% to 58.3% (relaxed) and from 37.1% to 43.4% (strict) within 50 frames, and in every setting more boundaries are recovered than lost.

Relaxed: at least one side aligned counts as positive

ToleranceBefore %Recovered%Lost %After %
Any distance88.97.43.193.2
≤10 frames19.515.112.622.1
≤20 frames33.118.215.136.2
≤50 frames52.117.611.458.3

Strict: full transition required

ToleranceBefore %Recovered %Lost %After %
Any distance49.617.910.357.2
≤10 frames15.212.310.017.5
≤20 frames25.015.411.928.5
≤50 frames37.116.610.343.4

Overall frame flips. In addition, we track frame-level prediction flips on Breakfast before and after adaptation. On average, 18.6% frames move from error to correct and 10.3% from correct to error, a net gain of 8.3 points that accounts for the accuracy rise from 65.4 to 73.7, with about 0.15% of frames labelled.

BreakfastAfter: correctAfter: error
Before: correct55.1%10.3%
Before: error18.6%16.0%

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.