OnlineTAS: An Online Baseline for Temporal Action Segmentation

Temporal context plays a significant role in temporal action segmentation. In an offline setting, the context is typically captured by the segmentation network after observing the entire sequence. However, capturing and using such context information in an online setting remains an under-explored problem. This work presents the an online framework for temporal action segmentation. At the core of the framework is an adaptive memory designed to accommodate dynamic changes in context over time, alongside a feature augmentation module that enhances the frames with the memory. In addition, we propose a post-processing approach to mitigate the severe over-segmentation in the online setting. On three common segmentation benchmarks, our approach achieves state-of-the-art performance.

Paper

References (47)

Scroll for more · 35 remaining

Similar papers

Peer review

Reviewer aGP14/10 · confidence 5/52024-07-08

Summary

This paper proposes an online baseline for temporal action segmentation. The method is built upon causal TCN and integrates a GRU, attention-based feature aggregation, and a memory bank. A heuristic online post-processing method is also proposed.

Strengths

1. I like the statement claiming the contribution as a baseline for action segmentation. It is a modest and accurate claim. The paper provides a reasonable online baseline for this task. Online action segmentation is also a meaningful task. 2. The proposed post-processing method is intuitive and effective. 3. The experiments and the ablation studies show good results.

Weaknesses

1. The major concern is about whether the method can achieve real-time inference speed. This is crucial for the online setting. Currently, no inference speed is reported, for example, the FPS including the feature extraction. Although the authors discussed this in the limitation section, this is a major weakness, given that the method is proposed as a baseline for the online setting. 2. Experiments on new benchmarks such as Assembly101 would be interesting. 3. I would recommend the authors remove the statement "we are the first to establish a fully supervised online framework for TAS". There are some concurrent works that became available before the NeurIPS submission deadline or slightly after that, e.g., O-TALC and Progress-Aware Online Action Segmentation.

Questions

NA

Rating

4

Confidence

5

Soundness

3

Presentation

4

Contribution

2

Limitations

The limitation has been discussed regarding the real-time inference. This is good.

Authorsrebuttal2024-08-12

We thank the reviewer's acknowledgment of our response to the weaknesses and their prompt feedback. To clarify, “online” refers to processing data incrementally as it becomes available. It does not necessarily imply zero latency (i.e., real-time processing). Many existing video understanding methods which claim to be (and are accepted) as online do not work in real-time. For instance, (semi-)online video segmentation methods such as GenVIS [1] operate at 6 FPS and DEVA [2] at 6.6 FPS; while online action detection methods like Testra [3] at 12 FPS, MAT [4] runs at 8.1 FPS, E2e-load [5] at 8.7 FPS, and MATR [6] at 6.0 FPS. Among these, [3,4,6] are not “real-time” because of their optical flow extraction. Additionally, we have included the speed of the CUDA version of the TV-L1 for optical flow calculation, using the same hardware configuration as we mentioned, in the updated table below. With the CUDA version, our apporach can achieve a real-time inference speed up to 33 FPS. We believe that utilizing the cuda version of the same TV-L1 algorithm will have a minimal impact on the overall performance. | Optical Flow Algorithm | Optical Flow Calculation | I3D - RGB | I3D - Optical Flow | Ours (I3D input) | Ours (raw RGB input) | |-|-|-|-|-|-| | | Time (ms) / FPS| Time (ms) / FPS | Time (ms) / FPS | Time (ms) / FPS | Time (ms) / FPS | | TV-L1 (cpu) | 166.5 / 6 | 9.3 / 107 | 11.2 / 89 | 4.2 / 238 | 191.2 / 5 | | TV-L1 (cuda) | 4.8 / 208 | 9.3 / 107 | 11.2 / 89 | 4.2 / 238 | 29.5 / 33 | We hope this clarification addresses the reviewer's concern regarding the inference speed of our approach. [1] A Generalized Framework for Video Instance Segmentation. CVPR 2023 [2] Tracking Anything with Decoupled Video Segmentation. ICCV 2023 [3] Real-time Online Video Detection with Temporal Smoothing Transformers. ECCV2022 [4] Memory-and-anticipation transformer for online action understanding. ICCV 2023 [5] E2e-load: End-to-end long-form online action detection. ICCV 2023 [6] Online Temporal Action Localization with Memory-Augmented Transformer. ECCV 2024

Reviewer itXK6/10 · confidence 4/52024-07-10

Summary

This paper introduces OnlineTAS, the first fully-supervised online framework for temporal action segmentation (TAS). The main contributions include: 1. A context-aware feature augmentation (CFA) module that incorporates an adaptive memory to enhance frame representations with temporal context. 2. An adaptive memory bank that accumulates short-term and long-term context information. 3. A post-processing technique for online boundary adjustment to mitigate over-segmentation. 4. State-of-the-art performance on three TAS benchmarks in an online setting.

Strengths

1. The CFA module and adaptive memory bank provide an innovative approach to incorporating temporal context in an online setting. 2. The proposed post-processing technique effectively mitigates the over-segmentation problem common in online TAS. 3. The authors demonstrate the effectiveness of their approach on three standard TAS benchmarks, showing consistent improvements over baselines.

Weaknesses

1. The paper lacks a thorough discussion of the computational complexity and resource requirements of OnlineTAS. For an online method intended for real-time applications, this is a critical aspect. The authors mention that their approach uses a "single-stage TCN" for efficiency, but don't provide concrete details on inference time, memory usage, or how these scale with video length. A comparison of computational requirements with offline methods and other online video understanding tasks would be beneficial. This analysis should include considerations for both the online and semi-online inference modes. 2. The paper presents quantitative results but lacks a qualitative error analysis. A detailed examination of failure cases could provide valuable insights into the limitations of the approach and guide future improvements. For instance, analyzing scenarios where OnlineTAS performs poorly compared to offline methods or where over-segmentation persists despite post-processing could be informative. 3. While the paper introduces an adaptive memory bank, there's limited discussion on how this approach scales to very long videos or continuous video streams. It's unclear how the method would perform in scenarios where the video length greatly exceeds the memory capacity, which is a likely scenario in many real-world applications. 4. For an online method, real-time performance is crucial. However, the paper doesn't provide metrics such as frames per second processing speed or end-to-end latency on standard hardware. This information is essential for assessing the practical applicability of the method in real-time scenarios.

Questions

The following questions are mostly related to weaknesses directly or indirectly: 1. Have you explored the impact of different feature extractors and different backbones on the performance of OnlineTAS? 2. Can you elaborate on how OnlineTAS might be adapted for weakly-supervised or unsupervised settings? 3. How sensitive is the method to the choice of hyperparameters, particularly the clip size and memory length? 4. Your post-processing technique seems effective in mitigating over-segmentation. Have you explored how this technique might be integrated into the model itself during training, rather than as a post-processing step? 5. The paper introduces both online and semi-online inference modes. How do these two modes compare in terms of latency and accuracy trade-offs? In what scenarios would one be preferable over the other?

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

The authors acknowledge some limitations, such as the computational intensity of the offline segmentation model used as a backbone and the reliance on pre-extracted features. These factors could hinder real-time application of the framework. However, the discussion of limitations could be expanded to include potential challenges in adapting the method to more diverse datasets or real-world streaming scenarios.

Reviewer xPL14/10 · confidence 5/52024-07-11

Summary

This paper presents the first online framework for temporal action segmentation. At the core of the framework is an adaptive memory designed to accommodate dynamic changes in context over time, alongside a feature augmentation module that enhances the frames with the memory. A post-processing approach is proposed to mitigate the severe over-segmentation in the online setting. The method achieves state-of-the-art performance on three common segmentation benchmarks.

Strengths

1. The paper is well written and easy to be followed. 2. Exploring online paradigms in temporal action segmentation is meaningful.

Weaknesses

1. Novelty is a big issue. The proposed adaptive memory bank mechanism has been explored in the video object segmentation tasks[1]. Also, semi-online inference scheme has been proposed in video instance segmentation tasks[3]. [1] Video Object Segmentation with Dynamic Memory Networks and Adaptive Object Alignment, iccv [2]Video Object Segmentation with Adaptive Feature Bank and Uncertain-Region Refinement,neurips [3] A Generalized Framework for Video Instance Segmentation, cvpr 2. The context-aware feature augmentation (CFA) module mainly consists of self-attention and cross-attention, adaptive memory bank was widely used in video detection and segmentation tasks, so the proposed module is not novel enough. 3. Is the Trans. Decoder necessary in the CFA module? For example, directly using memory as K and V of CA or performing SA together with ck. 4. In semi-online inference, why use non-overlapping clips? Would using overlapping clips for sampling and voting classification yield better results? 5. The performance improvement of Edit and F1 brought by post-processing is so amazing, why is there such a big improvement but Acc has decreased. 6. Compared to the latest offline methods, there is still a significant gap in the performance of online methods. The online method has advantages in terms of training cost and inference speed.

Questions

As an online paradigm, why use offline models as baselines instead of building online models from scratch? The solution of replacing standard convolutions with causal convolutions is not very elegant.

Rating

4

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

see weakness and questions.

Reviewer aGP12024-08-11

Response

Thanks for the rebuttal. W1: The method cannot achieve real-time speed using TV-L1. The method can achieve real-time speed using NVOFA, but it is unclear how using NVOFA affects the performance. I am satisfied with W2 and W3 at this stage. Overall, I tend to keep my score. But I am fine with acceptance if all others vote for accept.

Authorsrebuttal2024-08-13

Thanks for the comment

Dear reviewer itXK Thank you for taking the time to thoroughly review our rebuttal. We greatly appreciate your constructive feedback and are pleased that our responses have addressed your concerns.

Reviewer xPL12024-08-14

Thanks for your response. I have read the rebuttal carefully.

Authorsrebuttal2024-08-14

Thanks for the comment

Dear Reviewer xPL1, Thank you for taking the time to carefully read our rebuttal. If there are any aspects of our response that require further clarification or if you have any additional questions, we are more than happy to engage in further discussion.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC