Pith. sign in

REVIEW 4 minor 41 references

Sampling Gaussian frame embeddings before optimal-transport pseudo-labeling breaks the local-optimum trap that traps deterministic unsupervised action segmenters.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 07:27 UTC pith:35GK6FLM

load-bearing objection Clean, reproducible fix: Gaussian frame embeddings + GCN sampling before OT give solid gains over ASOT/VASOT without changing the OT math.

arxiv 2607.05263 v1 pith:35GK6FLM submitted 2026-07-06 cs.CV

Learning Probabilistic Embeddings for Unsupervised Action Segmentation

Shuai Li , Duc Manh Vu , Juergen Gall This is my paper
classification cs.CV
keywords unsupervised action segmentationprobabilistic embeddingsoptimal transportgraph convolutional networkstemporal action segmentationGaussian embeddings
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Unsupervised temporal action segmentation of long untrimmed videos usually alternates between learning a single deterministic vector for every frame and using optimal transport to turn those vectors into pseudo-labels. That loop quickly over-fits noisy labels and freezes. This paper replaces the deterministic vector with a Gaussian whose mean and variance are predicted by a graph-convolutional network; during training it draws samples from those Gaussians before each transport step. The extra randomness keeps the representation from locking onto early mistakes, so the same transport procedure yields substantially more accurate segmentations. On four standard benchmarks the change lifts mean-over-frames by as much as twenty percent and F1 by as much as nineteen percent relative to otherwise identical deterministic baselines, while inference cost stays unchanged because only the mean is used at test time.

Core claim

Modeling each frame embedding as a diagonal Gaussian and sampling from it before every optimal-transport pseudo-label step, rather than feeding a single deterministic vector into the transport, prevents the joint representation-and-clustering loop from collapsing into a poor local optimum and produces measurably better unsupervised action segmentations.

What carries the argument

Probabilistic embeddings: a GCN predicts per-frame mean and log-variance; the re-parameterization trick draws Monte-Carlo samples that become the cost matrix for fused Kantorovich–Gromov-Wasserstein optimal transport; the resulting soft assignments supervise an uncertainty-aware cross-entropy loss.

Load-bearing premise

The number of action classes for every activity is known in advance and set exactly to the ground-truth count; if that number must be discovered from data the reported gains may shrink.

What would settle it

Run the identical GCN-plus-sampling pipeline on the same four datasets while forcing K to be estimated by a standard model-selection criterion (e.g., silhouette or BIC on the learned means) instead of the ground-truth class count; if the MoF and F1 lifts over the deterministic baseline disappear, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Any existing OT-based unsupervised segmenter can be upgraded simply by swapping its deterministic encoder for a mean-and-variance GCN and sampling before transport.
  • The same sampling-before-clustering idea can be dropped into other joint-representation-and-clustering loops that currently over-fit early pseudo-labels.
  • Because only the mean is used at inference, the accuracy gains come with zero extra test-time cost.
  • The learned per-frame variance itself becomes a free uncertainty map that correlates with segmentation errors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same Gaussian-sampling trick may stabilize other EM-style video pipelines that alternate between assignment and representation learning, such as unsupervised action detection or multi-view clustering.
  • If the Monte-Carlo sample count is annealed with training progress (more samples early when uncertainty is high), further gains may appear without raising asymptotic cost.
  • The method’s reliance on a known K suggests a natural next experiment: replace the fixed action prototypes with a Dirichlet-process or Chinese-restaurant prior and measure how much of the reported lift survives.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

0 major / 4 minor

Summary. The paper proposes PEOT, which replaces deterministic frame embeddings in OT-based unsupervised temporal action segmentation with probabilistic embeddings modeled as diagonal Gaussians. A GCN predicts per-frame mean and log-variance; samples are drawn via the reparameterization trick before computing the fused Kantorovich–Gromov-Wasserstein OT pseudo-labels that supervise a cross-entropy loss (Eqs. 5–11). At inference only the mean is used. The method is evaluated on Breakfast, YTI, 50Salads (Eval) and Desktop Assembly, reporting MoF/F1/mIoU after activity-level Hungarian matching. Controlled ablations (Tables 2–3) show that grafting the same probabilistic layer onto ASOT and VASOT yields consistent gains (up to +12.2 MoF / +12.1 F1), and PEOT is best or second-best on 9 of 12 metrics.

Significance. If the reported gains hold, the work supplies a simple, drop-in improvement for the dominant joint-representation-and-OT paradigm. The ablations cleanly isolate the probabilistic component from the GCN architecture, multi-seed statistics appear in the supplement, and public code is provided. These elements make the contribution immediately usable by the community and strengthen the case that modeling embedding uncertainty mitigates early overfitting to noisy pseudo-labels.

minor comments (4)
  1. Abstract and several body paragraphs contain missing spaces (e.g., “proposetolearnaprobabilisticembedding”). A global pass for typesetting is needed.
  2. Sec. 4.1 states that K is set to the ground-truth number of actions per activity. While this is standard in the literature, a short remark on sensitivity to over-/under-estimated K would help readers who must choose K without labels.
  3. Fig. 2 caption and the surrounding text use both “covariance” and “variance”; the model predicts a diagonal covariance, so consistent terminology would avoid confusion.
  4. Table 1 reports literature numbers while Table 2 recomputes ASOT/VASOT with public code; a one-sentence note explaining the modest numerical differences would improve transparency.

Circularity Check

0 steps flagged

No significant circularity; purely empirical ML paper whose gains are measured against external ground-truth via Hungarian matching.

full rationale

The paper proposes replacing deterministic frame embeddings with Gaussians (sampled via reparameterization before OT pseudo-label estimation) inside an otherwise standard joint representation-learning + OT pipeline (ASOT/VASOT). All claims are empirical: MoF/F1/mIoU improvements are obtained by matching predicted segments to held-out ground-truth labels with the Hungarian algorithm on four public datasets. No equation, uniqueness theorem, or fitted constant is redefined as a “prediction”; hyperparameters follow the cited baselines and are ablated; self-citations are limited to ordinary prior methods used as controls. The known-K assumption is field-standard and does not create a definitional loop. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 1 invented entities

The work rests on the standard unsupervised action-segmentation experimental protocol (known K, Hungarian matching at activity level, pre-extracted IDT or similar features) plus the fused Kantorovich–Gromov-Wasserstein OT formulation taken from ASOT. The only new modeling choice is the diagonal-Gaussian embedding whose parameters are produced by a one-layer GCN; no free physical constants or invented particles appear.

free parameters (3)
  • number of Monte-Carlo samples M = 3
    Chosen by ablation (Table 7); M=3 is used for all main results. Affects both training cost and final accuracy.
  • GCN neighborhood size = 3 frames
    Fixed to 3-frame connectivity after ablation (Table 5); larger neighborhoods degrade performance via over-smoothing.
  • OT hyper-parameters (α, r, ρ, λ) = dataset-specific (e.g. α=0.4, r=0.04 on Breakfast)
    Taken from ASOT and lightly retuned per dataset (supplement sensitivity plots). Directly control the balance between visual cost and temporal consistency.
axioms (3)
  • domain assumption The number of action classes K equals the ground-truth count for each activity.
    Stated in Sec. 4.1; common in the literature but required for the reported numbers.
  • ad hoc to paper Frame features are well-modeled by independent diagonal Gaussians whose parameters can be predicted from a local temporal graph.
    Core modeling choice introduced in Sec. 3.2; no independent justification beyond empirical gains.
  • domain assumption The fused KOT+GWOT objective of ASOT is a suitable surrogate for temporal action segmentation.
    Adopted unchanged from Xu & Gould 2024; all gains are measured relative to this baseline.
invented entities (1)
  • PEOT probabilistic embedding layer (GCN-mean + GCN-var + reparameterized sampling) no independent evidence
    purpose: To inject data-dependent uncertainty into the frame representations so that OT pseudo-labels explore more assignments during training.
    New architectural component; independent evidence is limited to the ablation tables and qualitative self-similarity matrices inside this paper.

pith-pipeline@v1.1.0-grok45 · 20890 in / 2372 out tokens · 21217 ms · 2026-07-11T07:27:56.274051+00:00 · methodology

0 comments
read the original abstract

This paper concerns the problem of unsupervised temporal action segmentation for long, untrimmed videos. Recent successful approaches follow a joint representation learning and clustering paradigm, where optimal transport (OT) is adopted to produce pseudo labels for learning frame representations. These approaches alternate between estimating pseudo labels using OT and optimizing the parameters with gradient descent during training, where OT is used for obtaining the final temporal action segmentation. A major limitation of these works is that they learn a deterministic embedding for frame representations. The iterative procedure between learning deterministic embeddings based on pseudo labels and estimating pseudo labels from the learned embedding can thus get quickly stuck in a local optimum. As an alternative, we thus propose to learn a probabilistic embedding for frame representations. The embeddings are modeled by Gaussian distributions and we sample from the distributions before estimating the pseudo labels. We evaluate our approach on several challenging temporal action segmentation datasets and achieve results comparable to, and in some cases, better than the state of the art. Compared to baselines with deterministic embeddings, our approach improves MoF up to 20.7\% and F1-score up to 19.0\%. Our code is available at https://github.com/derkbreeze/PEOT.

Figures

Figures reproduced from arXiv: 2607.05263 by Duc Manh Vu, Juergen Gall, Shuai Li.

Figure 1
Figure 1. Figure 1: Most previous works learn deterministic embeddings as frame embeddings; we propose to learn probabilistic embeddings such that frame representations are samples from Gaussian distributions that explicitly model embedding uncertainty. Our repre￾sentations lead to more accurate segmentations compared to the baseline [38]. Different colors indicate different actions. loss. Following this paradigm, Xu and Goul… view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of our training scheme. Features of an input video are fed through an MLP to obtain per-frame embeddings X. A temporally weighted graph is constructed, and the normalized adjacency matrix A˜ along with frame embeddings X are fed into a Graph Convolutional Network (GCN), which produces the mean and covariance of the frame embeddings. The re-parametrization trick is used during training to obtain pr… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results. Comparing ASOT [38], PEOT (ours), and ground truth (GT) across different datasets and activities [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results. Two examples when using VASOT [2] as baseline. The examples are from the Breakfast [14] dataset. 4.4 Qualitative Results We finally provide some qualitative segmentation results in [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results of the learned frame-wise representations. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Analysis of the learned uncertainty. GT Epoch02 Epoch14 PEOT w.o. Prob PEOT [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Difference between probabilistic embeddings and deterministic embeddings dur￾ing training. After 2 epochs, the approach using the deterministic embedding (row 2) recognizes the pink action, while the probabilistic embedding (row 3) misses it. After 14 epochs, the segmentation with the deterministic embedding only changes by shift￾ing the boundaries between the segments whereas the pink action now also appe… view at source ↗
Figure 1
Figure 1. Figure 1: Comparing for ASOT [38] (dashed) and our approach PEOT (solid) the loss curves (black) and evaluation metrics MoF (red) and F1 (blue) for videos of the friedegg and changing_tire activity from the Breakfast [14] and Youtube Instr. [1] dataset, respectively. We scale the loss for better visualization. B Sensitivity Analysis For the parameters of the fused Gromov-Wasserstein optimal transport, we fol￾low ASO… view at source ↗
Figure 2
Figure 2. Figure 2: Sensitivity analysis of OT hyper-parameters. C Training Time We use the same OT formulation as ASOT [38] and conducted all experiments on a single NVIDIA GeForce RTX 3090 GPU (24GB) with CUDA 11.8. Compared to ASOT, our method PEOT only contains one additional GCN layer followed by a sampling procedure during training. The results in [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 5 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Alayrac, J.B., Bojanowski, P., Agrawal, N., Sivic, J., Laptev, I., Lacoste-Julien, S.: Unsupervised learning from narrated instruction videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4575–4583 (2016) 16 S. Li et al

    2016

  2. [2]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Ali, A.S., Mahmood, S.A., Saeed, M., Konin, A., Zia, M.Z., Tran, Q.H.: Joint self-supervised video alignment and action segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10807–10818 (2025)

    2025

  3. [3]

    In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision

    Bueno-Benito, E., Dimiccoli, M.: Clot: Closed loop optimal transport for unsuper- vised action segmentation. In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision. pp. 10719–10729 (2025)

    2025

  4. [4]

    Mathematics of computation87(314), 2563– 2609 (2018)

    Chizat, L., Peyré, G., Schmitzer, B., Vialard, F.X.: Scaling algorithms for unbal- anced optimal transport problems. Mathematics of computation87(314), 2563– 2609 (2018)

    2018

  5. [5]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    De Plaen, H., De Plaen, P.F., Suykens, J.A., Proesmans, M., Tuytelaars, T., Van Gool, L.: Unbalanced optimal transport: A unified framework for object de- tection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3198–3207 (2023)

    2023

  6. [6]

    Journal ofthe royalstatistical society: seriesB (method- ological)39(1), 1–22 (1977)

    Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data viathe emalgorithm. Journal ofthe royalstatistical society: seriesB (method- ological)39(1), 1–22 (1977)

    1977

  7. [7]

    In: Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition

    Farha, Y.A., Gall, J.: Ms-tcn: Multi-stage temporal convolutional network for ac- tion segmentation. In: Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition. pp. 3575–3584 (2019)

    2019

  8. [8]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Huang, Y., Sugano, Y., Sato, Y.: Improving action segmentation via graph-based temporal reasoning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14024–14034 (2020)

    2020

  9. [9]

    Kendall, A., Gal, Y.: What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems30(2017)

    2017

  10. [10]

    In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Khan, H., Haresh, S., Ahmed, A., Siddiqui, S., Konin, A., Zia, M.Z., Tran, Q.H.: Timestamp-supervised action segmentation with graph convolutional networks. In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 10619–10626. IEEE (2022)

    2022

  11. [11]

    arXiv preprint arXiv:1412.6980 (2014)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  12. [12]

    arXiv preprint arXiv:1312.6114 (2013)

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)

    Pith/arXiv arXiv 2013

  13. [13]

    arXiv preprint arXiv:1609.02907 (2016)

    Kipf, T.: Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

    Pith/arXiv arXiv 2016

  14. [14]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Kuehne, H., Arslan, A., Serre, T.: The language of actions: Recovering the syn- tax and semantics of goal-directed human activities. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 780–787 (2014)

    2014

  15. [15]

    Naval research logistics quarterly2(1-2), 83–97 (1955)

    Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly2(1-2), 83–97 (1955)

    1955

  16. [16]

    In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition

    Kukleva, A., Kuehne, H., Sener, F., Gall, J.: Unsupervised learning of action classes with continuous temporal embedding. In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition. pp. 12066–12074 (2019)

    2019

  17. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Kumar, S., Haresh, S., Ahmed, A., Konin, A., Zia, M.Z., Tran, Q.H.: Unsuper- vised action segmentation by joint representation learning and online clustering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20174–20185 (2022)

    2022

  18. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Li, J., Todorovic, S.: Action shuffle alternating learning for unsupervised action segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12628–12636 (2021) Learning Probabilistic Embeddings for Unsupervised Action Segmentation 17

    2021

  19. [19]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Li, S., Kong, Y., Rezatofighi, H.: Learning of global objective for network flow in multi-object tracking. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8855–8865 (2022)

    2022

  20. [20]

    Advances in Neural Information Processing Systems36, 18046–18075 (2023)

    Nguyen, K., Ho, N.: Energy-based sliced wasserstein distance. Advances in Neural Information Processing Systems36, 18046–18075 (2023)

    2023

  21. [21]

    arXiv preprint arXiv:1810.00319 (2018)

    Oh, S.J., Murphy, K., Pan, J., Roth, J., Schroff, F., Gallagher, A.: Modeling uncer- tainty with hedged instance embedding. arXiv preprint arXiv:1810.00319 (2018)

    Pith/arXiv arXiv 2018

  22. [22]

    In: International conference on machine learning

    Peyré, G., Cuturi, M., Solomon, J.: Gromov-wasserstein averaging of kernel and distance matrices. In: International conference on machine learning. pp. 2664–2672. PMLR (2016)

    2016

  23. [23]

    Engineering Applications of Artificial Intelli- gence148(2025)

    Romeo, L., Marani, R., Perri, A.G., Gall, J.: Multi-modal temporal action segmen- tation for manufacturing scenarios. Engineering Applications of Artificial Intelli- gence148(2025)

    2025

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Sarfraz, S., Murray, N., Sharma, V., Diba, A., Van Gool, L., Stiefelhagen, R.: Temporally-weighted hierarchical clustering for unsupervised action segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11225–11234 (2021)

    2021

  25. [25]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Sener, F., Yao, A.: Unsupervised learning and segmentation of complex activities from video. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8368–8376 (2018)

    2018

  26. [26]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Shi, Y., Jain, A.K.: Probabilistic face embeddings. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6902–6911 (2019)

    2019

  27. [27]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Spurio, F., Bahrami, E., Francesca, G., Gall, J.: Hierarchical vector quantization for unsupervised action segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 6996–7005 (2025)

    2025

  28. [28]

    In: Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing

    Stein, S., McKenna, S.J.: Combining embedded accelerometers with computer vi- sion for recognizing food preparation activities. In: Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing. pp. 729–738 (2013)

    2013

  29. [29]

    In: European Conference on Computer Vision

    Sun, J.J., Zhao, J., Chen, L.C., Schroff, F., Adam, H., Liu, T.: View-invariant probabilistic embedding for human pose. In: European Conference on Computer Vision. pp. 53–70. Springer (2020)

    2020

  30. [30]

    In: 2021 IEEE International Conference on Image Processing (ICIP)

    Swetha, S., Kuehne, H., Rawat, Y.S., Shah, M.: Unsupervised discriminative em- bedding for sub-action learning in complex activities. In: 2021 IEEE International Conference on Image Processing (ICIP). pp. 2588–2592. IEEE (2021)

    2021

  31. [31]

    Notes of Course at University of Cambridge3(2018)

    Thorpe, M.: Introduction to optimal transport. Notes of Course at University of Cambridge3(2018)

    2018

  32. [32]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Tran, Q.H., Mehmood, A., Ahmed, M., Naufil, M., Zafar, A., Konin, A., Zia, Z.: Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6426–6436 (2024)

    2024

  33. [33]

    Advances in neural information pro- cessing systems30(2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017)

    2017

  34. [34]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    VidalMata, R.G., Scheirer, W.J., Kukleva, A., Cox, D., Kuehne, H.: Joint visual- temporal embedding for unsupervised learning of actions in untrimmed sequences. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1238–1247 (2021)

    2021

  35. [35]

    arXiv preprint arXiv:1412.6623 (2014) 18 S

    Vilnis, L., McCallum, A.: Word representations via gaussian embedding. arXiv preprint arXiv:1412.6623 (2014) 18 S. Li et al

    Pith/arXiv arXiv 2014

  36. [36]

    In: Pro- ceedings of the IEEE international conference on computer vision

    Wang, H., Schmid, C.: Action recognition with improved trajectories. In: Pro- ceedings of the IEEE international conference on computer vision. pp. 3551–3558 (2013)

    2013

  37. [37]

    In: Proceedings of the European conference on computer vision (ECCV)

    Wang, X., Gupta, A.: Videos as space-time region graphs. In: Proceedings of the European conference on computer vision (ECCV). pp. 399–417 (2018)

    2018

  38. [38]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Xu, M., Gould, S.: Temporally consistent unbalanced optimal transport for un- supervised action segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14618–14627 (2024)

    2024

  39. [39]

    watching

    Yang, Y., Li, Y., Fermuller, C., Aloimonos, Y.: Robot learning manipulation action plans by “watching” unconstrained videos from the world wide web. In: Proceedings of the AAAI conference on artificial intelligence. vol. 29 (2015)

    2015

  40. [40]

    In: European Conference on Computer Vision

    Zatsarynna, O., Bahrami, E., Farha, Y.A., Francesca, G., Gall, J.: Gated temporal diffusion for stochastic long-term dense anticipation. In: European Conference on Computer Vision. pp. 454–472. Springer (2024)

    2024

  41. [41]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Zeng, R., Huang, W., Tan, M., Rong, Y., Zhao, P., Huang, J., Gan, C.: Graph convolutional networks for temporal action localization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7094–7103 (2019) Learning Probabilistic Embeddings for Unsupervised Action Segmentation Supplementary Materials Shuai Li1, Duc Manh Vu1, and Ju...

    2019