acceptodds
Withdrawn from ICLR 2027

Video-XAI Bench: A Known-Evidence Benchmark for Spatiotemporal Explanation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Video classifiers flag events for human review in public-safety and security monitoring. A classifier's decision is hard to trust unless a reviewer can check the evidence behind it. Explanation methods are meant to provide this evidence by highlighting the regions and moments of a video that drove the decision. Existing evaluations check if explanations look plausible or if removing the highlighted content lowers the model's confidence; they do not test if explanations point to the right place and moment. We introduce Video-XAI Bench, a benchmark for explanations of video classifiers. It tests explanation methods in two settings. The first uses real violence-recognition videos. The second adds a moving pattern to real videos at a known place and time, so we know exactly where and when the evidence is. On real videos, no method consistently performs best, as the best choice changes with the dataset, the model, and the score. With known evidence, scores based on the model's confidence do not reveal if an explanation finds the right moment, and some of the explanations that find it best partly reflect the input video rather than what the model learned. In a pilot with 18 participants, those who agreed that a video showed aggression often disagreed about which frames showed it, so human-marked evidence on real videos forms a window with soft edges. Our study concludes that video explanations should be tested against known evidence, not only against the model's confidence.

Primary area:
interpretability and explainable AI
Keywords:
Explainable AI, Video Explainability, Post-hoc Explanations, Trustworthy AI

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_PYN5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a benchmark for evaluating spatiotemporal explanations of video classifiers. The authors compare explanation methods using faithfulness, localization, validity, and runtime measures, and argue that confidence-based perturbation scores do not reliably reveal whether an explanation points to the actual evidence.

Critical strengths
  1. The known-evidence setting is a clean way to expose weaknesses in popular explanation metrics.

  2. Separating temporal localization from confidence change is useful, and the controls are more thoughtful than a simple visual plausibility comparison.

Critical weaknesses
  1. The benchmark establishes that methods can locate an artificial cue, but it does not establish that they explain a classifier's meaningful video reasoning. A moving grating is deliberately easy to define and may be detected through low-level texture or frequency statistics. That is a different task from identifying the action, interaction, or temporal relation that a real violence classifier uses. The manuscript's public-safety motivation therefore reaches well beyond what the benchmark validates.

  2. There is also a conceptual issue with the main evaluation story. If deletion and insertion scores can disagree with known localization, the strongest conclusion may be that these metrics lack semantic validity. Ranking explanation methods by those same metrics then becomes difficult to interpret. The real-video experiments do not solve this problem because the human evidence labels are soft and the datasets are small. At present, the paper is a valuable diagnostic study, but the claim that it provides a general benchmark for video explanations is too broad.

Questions for authors
  1. Do the conclusions survive several cue types, including motion, object interaction, and changes in shape, scale, and speed?
  2. What evidence supports transfer from the controlled cue to natural video explanations?
Reviewer_BiarOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Video-XAI Bench for evaluating spatiotemporal explanations of video classifiers. The real video study compares nine explanation methods across five architectures on three datasets. The controlled study inserts moving patterns with known spatial support and temporal support. Evaluation measures confidence based faithfulness, temporal localization, model dependence. The results expose discrepancies between these criteria. A pilot study with 18 participants examines agreement on evidence frames in five real videos.

Critical strengths
  1. The paper provides a useful diagnostic of explanation metrics. The oracle comparisons show that high insertion scores can coexist with poor temporal selectivity. This finding gives concrete support for evaluating temporal localization separately from changes in model confidence.

  2. The controlled study includes substantive validation. Matched static patterns, shuffled phases, targeted masking, single frame probes test whether the classification task requires the intended evidence. Experiments use three training seeds. Statistical inference treats source videos as independent units. Randomization checks, reference maps, method roster ablations strengthen the analysis within the tested setting.

Critical weaknesses
  1. The fixed temporal support limits the validity of the localization benchmark. Every controlled sample places the cue in frames 4 through 27. A fixed map concentrating on this interval obtains perfect TLS without locating evidence in the current video. The reported randomization checks do not distinguish a learned positional prior from localization based on video content. This limitation affects the central claim that the benchmark evaluates temporal evidence localization.

  2. The main findings depend on explanation configurations whose effects remain unresolved. The controlled IG pipeline includes absolute channel aggregation, smoothing, percentile clipping, power sharpening. The controlled Occlusion pipeline masks all frames together, making temporal localization unselective by construction. The roster ablations document substantial sensitivity to these methods. The paper does not isolate the contribution of IG postprocessing or evaluate temporally resolved Occlusion in the controlled setting. These omissions limit conclusions about the explanation methods beyond the specific configurations tested.

Questions for authors
  1. Experimental validation: Does the temporal localization result persist when cue onset varies across samples? Can the authors test whether explanations follow a translated cue while the classifier remains accurate? This would help distinguish content based localization from a fixed positional prior.

  2. Configuration validation: Do the main findings persist with IG before the additional postprocessing? What changes when Occlusion uses a sliding temporal window in the controlled setting? These comparisons would clarify whether the observed metric discrepancies depend on the selected implementations.

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

This papar investigates whether we correctly evaluate the explanations obtained from XAI algorithms. Authors claim that reperesentative protocol such as insertion and deletion does not ensure, precise indicating spatio-temporal region where the evidences exist. To test this, Video-XAI Bench combines real-video evaluation across three violence-recognition datasets and controlled task with synthetic motion cues. This controlled setup evaluates faithfulness, localization and model dependence seperately.

Critical strengths
  1. Authors employed careful controls to validate the synthetic task such as motion conditions, target masking, single-frame probes, and other tests to verify. that the whether classifier rely on the temporal cue or trivial apperance or background.
Critical weaknesses
  1. Fundamental limitations in the construct validity of temporal explanation evaluation. All controlled videos contain the cue in the fixed temporal window (frames 4-27). This potentailly allow an input-independent temporal prior to achieve high (even perfect) TLS. More fundamentally, the localization metric measure whether sailency is in the cue's support, rather than measure the temporal relationship. For example, if a model's sailency falls in to single pixel within the GT region can achieve perfect score. Consequently, the proposed benchmark does not adequately measure the correctness of temporal explanations.

  2. Weak connection between the real-video and controlled evaluations. The real-video setup evaluates violence recognition, whereas the controlled setup evaluates a synthetic motion coherence task. Moreover, several XAI methods use different configurations across these two setups. Although authors acknowledge these differences, no experiment establishes whether the failures observed in the controlled setting also occur in real-video classification.

Questions for authors
  1. Could evaluation protocol be improved? For example, random temporal window and avoiding the evaluation shortcuts I mentioned above weakness 1.

  2. Could you specify the connection between the two evaluation setups? (real and controlled)

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

Related papers

Discussion (0)

Sign in to comment.