CoChat

MS-TCN++: Multi-Stage Temporal Convolutional Network for Action Segmentation

Li, Shijie, Farha, Yazan Abu, Liu, Yun, Cheng, Ming-Ming, Gall, Juergen · 2023 · IEEE Transactions on Pattern Analysis and Machine Intelligence

The DOI matches its Crossref record, with a small difference in title or year; 56 of the 60 references it cites were found in a registry or an index; 4 of the 14 sources its claims cite were checked for support.

Abstract

With the success of deep learning in classifying short trimmed videos, more attention has been focused on temporally segmenting and classifying activities in long untrimmed videos. State-of-the-art approaches for action segmentation utilize several layers of temporal convolution and temporal pooling. Despite the capabilities of these approaches in capturing temporal dependencies, their predictions suffer from over-segmentation errors. In this paper, we propose a multi-stage architecture for the temporal action segmentation task that overcomes the limitations of the previous approaches. The first stage generates an initial prediction that is refined by the next ones. In each stage we stack several layers of dilated temporal convolutions covering a large receptive field with few parameters. While this architecture already performs well, lower layers still suffer from a small receptive field. To address this limitation, we propose a dual dilated layer that combines both large and small receptive fields. We further decouple the design of the first stage from the refining stages to address the different requirements of these stages. Extensive evaluation shows the effectiveness of the proposed model in capturing long-range dependencies and recognizing action segments. Our models achieve state-of-the-art results on three datasets: 50Salads, Georgia Tech Egocentric Activities (GTEA), and the Breakfast dataset.

Overview

CoChat Summary

TL;DR

The paper introduces MS-TCN, a multi-stage temporal convolutional network for fully supervised temporal action segmentation that operates at full temporal resolution and refines predictions stage by stage, trained with a combined cross-entropy and truncated-MSE smoothing loss. It reports state-of-the-art results on 50Salads, GTEA, and Breakfast.

Key findings

  • The multi-stage architecture substantially outperforms a single-stage TCN of equal parameter count (48 layers) on 50Salads: F1@{10,25,50} of 76.3/74.0/64.5 vs 49.0/46.4/40.2, Edit 67.9 vs 40.7, Acc 80.7 vs 78.0 — a margin of up to 27%.
  • The proposed smoothing loss improves over cross-entropy alone on 50Salads: L_cls gives F1@{10,25,50} 71.3/69.7/60.7, Edit 64.2, Acc 79.9, while L_cls + λL_T-MSE gives 76.3/74.0/64.5, Edit 67.9, Acc 80.7 (an absolute F1/edit improvement of ~5%); it also beats a KL-divergence smoothing loss (71.9/69.3/60.1, Edit 64.6, Acc 80.2).
  • Four stages is optimal; adding a fifth degrades performance (F1@10 76.4 but F1@25/50 drop to 73.4/63.6 and Acc to 79.5), attributed to over-fitting from more parameters.
  • Concatenating features to the input probabilities of higher stages causes a large drop (F1@10 56.2 vs 76.3, Edit 47.6 vs 67.9, ~20% drop), so only probabilities are passed.
  • The model outperforms state-of-the-art methods on all three datasets, with up to 12.6% improvement in frame-wise accuracy on 50Salads; on GTEA with fine-tuning it reaches F1@{10,25,50} 87.5/85.4/74.6, Edit 81.4, Acc 79.2.

Methods

  • Architecture: a single-stage TCN of dilated 1D convolutions (kernel size 3, acausal, dilation doubled per layer 1–512, residual connections, no pooling or fully connected layers), stacked into a multi-stage model where each stage refines the previous stage's frame-wise probabilities; implementation uses 4 stages, 10 dilated layers per stage, 64 filters, dropout, Adam with learning rate 0.0005.
  • Loss: sum over stages of a per-stage loss combining cross-entropy classification loss and a truncated mean squared error smoothing loss over frame-wise log-probabilities, with τ = 4 and λ = 0.15.
  • Data and evaluation: I3D features at 15 fps on 50Salads (50 videos, 17 classes, 5-fold cross-validation), GTEA (28 videos, 11 classes, leave-one-subject-out), and Breakfast (1,712 videos, 48 actions, standard 4 splits); metrics are frame-wise accuracy, segmental edit distance, and segmental F1 at IoU thresholds 10%, 25%, 50%.

Limitations

  • Stated: the weakly supervised setting is explicitly out of scope ("the weakly supervised case is beyond the scope of this paper"); performance is slightly worse on longer videos due to the limited receptive field (GTEA ≥1.5 min: F1@10 81.2, Edit 71.8, Acc 75.9 vs <1 min: 89.6, 82.5, 76.6); a fifth stage degrades results, possibly from over-fitting.
  • Evident from the text: evaluation is confined to cooking/egocentric activity datasets (50Salads, GTEA, Breakfast), so generalization to other domains is untested; the model is fully supervised and requires frame-level annotations.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/tpami.2020.3021756

Cites

56 of 56 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Overview of the multi-stage temporal convolutional network. Each stage generates an initial prediction that is refined by the next stage. At each stage, several dilated 1D convolutions are applied on the activations of the previous layer. A loss layer is added after each stage.
Figure 1 Overview of the multi-stage temporal convolutional network. Each stage generates an initial prediction that is refined by the next stage. At each stage, several dilated 1D convolutions are applied on the activations of the previous layer. A loss layer is added after each stage.From the PDF
Figure 2 Overview of the dilated residual layer.
Figure 2 Overview of the dilated residual layer.From the PDF
Figure 3 Qualitative result from the 50Salads dataset for comparing different number of stages.
Figure 3 Qualitative result from the 50Salads dataset for comparing different number of stages.From the PDF
Figure 4 Qualitative result from the 50Salads dataset for comparing different loss functions.
Figure 4 Qualitative result from the 50Salads dataset for comparing different loss functions.From the PDF
Figure 5 Loss surface for the Kullback-Leibler (KL) divergence loss ( L KL ) and the proposed truncated mean squared loss ( L T -MSE ) for the case of two classes. y t,c is the predicted probability for class c and y t -1 ,c is the target probability corresponding to that class.
Figure 5 Loss surface for the Kullback-Leibler (KL) divergence loss ( L KL ) and the proposed truncated mean squared loss ( L T -MSE ) for the case of two classes. y t,c is the predicted probability for class c and y t -1 ,c is the target probability corresponding to that class.From the PDF
Figure 6 Qualitative results for two videos from the 50Salads dataset for showing the effect of passing features to higher stages.
Figure 6 Qualitative results for two videos from the 50Salads dataset for showing the effect of passing features to higher stages.From the PDF
Figure 7 Qualitative results for the temporal action segmentation task on (a) 50Salads (b) GTEA, and (c) Breakfast dataset.
Figure 7 Qualitative results for the temporal action segmentation task on (a) 50Salads (b) GTEA, and (c) Breakfast dataset.From the PDF

Verification

Claims the paper makes

3 Valid · 1 Not valid · 5 Not checked · 71 Own results and interpretations

45 more claims

Create a free account to see all 80 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 8
  • Borderline: one of the tests for a substantial claim is unclear. 1
  • It may state two separate claims as one. 22

Create a free account to see these claims.

Citations

Cites

Show all 56 references

Research works printed but not linked: 1

  1. karaman, 2014, Fast saliency based pooling of fisher encoded dense trajectories, Proc Eur Conf Comput Vis THUMOS Workshop

Cited by

See all 349

Metadata

Authors
Li, S., Farha, Y. A., Liu, Y., Cheng, M.-M., & Gall, J.
Affiliations
  • University of Bonn (DE)
  • Nankai University (CN)
  • University of Bonn, Bonn, Germany
Venue
IEEE Transactions on Pattern Analysis and Machine Intelligence
Published
2023-06-01
Volume
45
Issue
6
Pages
6647-6658
ISSN
  • 0162-8828 (ISSN-L)
  • 2160-9292
  • 1939-3539
OpenAlex
W3083550439
Cited by
349
References
60
Citations by year
2026: 45 · 2025: 69 · 2024: 74 · 2023: 49
FWCI
15.68
Type
journal article
Language
en
Open access
closed
Locations
  • IEEE Transactions on Pattern Analysis and Machine Intelligence (publishedVersion)
  • PubMed (publishedVersion)
Funders
  • Deutsche Forschungsgemeinschaft (GA 1927/4-1)
  • ERC Starting (677650)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Human Pose and Action Recognition
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Video Surveillance and Tracking Methods
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Vision and Imaging
Keywords
temporal action segmentation, long-range temporal dependencies, over-segmentation, untrimmed video, activity recognition, receptive fields

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: MS-TCN++: Multi-Stage Temporal Convolutional Network for Action Segmentation. https://app.cochat.ai/papers/10.1109/tpami.2020.3021756
The paper
Li, S., Farha, Y. A., Liu, Y., Cheng, M.-M., & Gall, J. (2023). MS-TCN++: Multi-Stage Temporal Convolutional Network for Action Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence. https://doi.org/10.1109/tpami.2020.3021756

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.