Pith. sign in

REVIEW 3 major objections 5 minor 64 references

Temporal Rate Reduction Clustering for Human Motion Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read TR2C jointly learns temporally consistent, union-of-subspace-aligned representations and a doubly stochastic affinity, reporting 94.12% accuracy on Weizmann and 97.96% on YouTube for unsupervised human motion segmentation.

desk verdict TR2C is a solid, empirically strong method paper — the new combination of rate reduction with a temporal Laplacian works on five HMS benchmarks — but the paper overclaims the theoretical grounding for its UoS-alignment story, and the per-dataset tuning is under-specified. read the letter →

arxiv 2506.21249 v2 pith:ROXIQZNT submitted 2025-06-26 cs.CV

Xianghan Meng , Zhengyu Tong , Zhiyuan Huang , Chun-Guang Li This is my paper · ORCID
classification cs.CV
keywords humanmotionsegmentationsubspaceclusteringmaximalcodingratereductiontemporalconsistencyrepresentationlearningunsupervisedvideospectralunionofsubspaces
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces TR2C, an unsupervised method for human motion segmentation that learns the frame representation and the clustering affinity at the same time. Its goal is to fix the main bottleneck of subspace-clustering approaches to this task: raw video features often do not lie on a union of subspaces, so segmenting them directly works poorly. TR2C adds a temporal smoothness term to a rate-reduction objective, which forces the learned representations to be temporally consistent and to concentrate into approximately orthogonal subspaces. The paper reports top clustering accuracy on five HMS benchmarks, with particularly large gains over transfer-learning baselines even though TR2C uses no additional source data. If correct, the method shows that jointly reshaping the data distribution and exploiting temporal order, rather than adding supervision or transfer, is the key to better motion segmentation.

What carries the argument

The load-bearing mechanism is the rate-reduction objective (Eq. 9) with its two coding-rate terms: $\rho(Z,\epsilon)$ measures the total volume of the representations and $\rho_c(Z,\epsilon\mid\Pi)$ measures the volume inside each cluster, so maximizing their difference inflates the global spread while compressing each cluster, driving the data into a union of roughly orthogonal subspaces. The temporal Laplacian $r(Z)=\operatorname{tr}(ZLZ^{\top})$ encodes the prior that neighboring frames belong to the same motion. The continuous relaxation replaces the hard partition $\Pi$ with a doubly stochastic affinity $\Gamma$ produced by a differentiable Sinkhorn projection of normalized cluster-head outputs, so the whole pipeline is trainable end-to-end by backpropagation. A final identity, $\log\det(I+ZZ^{\top})=\log\det(I+Z^{\top}Z)$, reduces the loss complexity from $O(N^4)$ to $O(Nd^3)$ and makes the method practical on long videos.

What would settle it

Train TR2C on a synthetic video sequence whose frames are drawn from known ground-truth subspaces, one per motion, with controlled noise, then measure the principal angles between the cluster means of the learned representation $Z$: the central claim predicts near-orthogonality. If the angles stay far from orthogonal while clustering accuracy on the planted labels is near chance, the alignment guarantee has not transferred to the relaxed objective; alternatively, if replacing the Sinkhorn affinity $\Gamma$ with a hard-assignment matrix from standard spectral clustering leaves performance unchanged, the affinity learning is not doing the claimed work.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that the HMS task can be formulated as $\min_{Z,\Pi}\, -\rho(Z,\epsilon)+\lambda_1\,\rho_c(Z,\epsilon\mid\Pi)+\lambda_2\,r(Z)$ subject to unit-norm columns (Eq. 9), where $\rho$ is the total coding rate, $\rho_c$ is the class-conditional coding rate under the partition $\Pi$, and $r(Z)=\operatorname{tr}(ZLZ^{\top})$ is a temporal Laplacian regularizer. Solving this problem jointly with neural reparameterizations—the feature head $Z=g(f(X))$, the cluster head $Y=h(f(X))$, and the affinity $\Gamma=\mathcal{P}_{\Xi}(\tilde{Y}^{\top}\tilde{Y})$ obtained by Sinkhorn projection—yields representations that the paper shows align with a union of approximately orthogonal subspaces and stay aligned under additive noise. On five HMS datasets, spectral clustering on $\Gamma$ attains 94.12% ACC on Weizmann, 83.50% on Keck, 93.54% on UT, 83.08% on MAD, and 97.96% on YouTube with HoG or VGG features, and 96.21%, 90.10%, and 99.35% on Weizmann, Keck, and YouTube when the input features come from a pretrained CLIP encoder. The paper reads these results as evidence that jointly learning structured representations and affinity, rather than improving the clustering step on fixed features, is what makes unsupervised temporal segmentation work.

Load-bearing premise

The paper assumes that the rate-reduction guarantee of union-of-subspaces alignment, proven for the supervised hard-partition problem, still holds after the partition is replaced by a learned doubly stochastic affinity and after both the representations and the affinity are generated by neural networks trained with backpropagation.

Editorial extensions

If this is right

  • Unsupervised HMS on raw HoG/VGG features reaches 94.12%, 83.50%, 93.54%, 83.08%, and 97.96% ACC on Weizmann, Keck, UT, MAD, and YouTube, surpassing transfer-learning baselines that use additional source data, according to the paper's comparison tables.
  • With pretrained CLIP features, TR2C reaches 96.21%, 90.10%, and 99.35% ACC on Weizmann, Keck, and YouTube, while the paper's zero-shot CLIP baseline only reaches 29.14% on Weizmann, indicating the temporal-context learning, not the feature alone, is responsible for the gains.
  • Ablations that remove any one of the three loss terms drop accuracy sharply (on Weizmann, from about 94% to 37-65% ACC), so the total rate, the conditional rate, and the temporal smoothness term are each load-bearing.
  • TR2C's learned representations degrade less under additive Gaussian noise than GCTSC's representations or raw HoG features (at most about 15% accuracy drop on Weizmann with EnSC), supporting the claim that union-of-subspaces alignment brings robustness.
  • The $\log\det$ commutation identity reduces the loss to $O(Nd^3)$, making TR2C more than 100x faster than GCTSC on the Weizmann sequence on a single GPU (10.7 s versus 1122.2 s).

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves open whether the MCR2 alignment guarantee, proven for the supervised hard-partition problem, survives the relaxation to a Sinkhorn-normalized affinity and neural reparameterization; a synthetic union-of-subspaces experiment with known ground-truth subspaces would settle this.
  • If the alignment mechanism is as general as the results suggest, the same objective could transfer to other sequential clustering problems, such as unsupervised audio event segmentation, biosignal segmentation, or time-series motif discovery, where temporal smoothness is a natural prior.
  • The per-dataset tuning of $\lambda_1$ (0.1-1.0) and $\lambda_2$ (2-12) suggests the balance between subspace compression and temporal smoothing is data-dependent, so a principled schedule for these weights could make the method parameter-free.
  • The paper's own error analysis on Keck and MAD points to motion-transition frames as the main failure locations, which implies that a boundary-resolving post-processing step could extend the reported gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Temporal Rate Reduction Clustering (TR2C) for unsupervised human motion segmentation. The method learns a network-parameterized embedding Z and a doubly stochastic affinity Γ by minimizing L = −Lρ + λ1 Lρc + λ2 Lr, where Lρ is the total coding rate, Lρc is the conditional coding rate computed through a Sinkhorn-projected affinity, and Lr is a temporal Laplacian smoothness term (Eqs. (9)–(14)). The learned affinity is then fed to spectral clustering. The paper reports ACC/NMI on Weizmann, Keck, UT, MAD, and YouTube with HoG/VGG features and with CLIP features, claiming state-of-the-art results (e.g., 94.12 ACC on Weizmann and 97.96 on YouTube in Table 2; 96.21 and 99.35 with CLIP in Table 5), together with ablations, hyper-parameter sensitivity, robustness to noise, runtime comparisons, and additional temporal action segmentation experiments. The central claim is that the learned representations are temporally consistent and aligned with a union of subspaces, which is argued to be the mechanism behind the accuracy gains.

Significance. If the central claims hold, the paper makes a useful contribution: it is the first application of the MCR2 principle to temporal clustering, it ships code, and the benchmark improvements over the strongest prior method GCTSC are large on several datasets (e.g., +9.1 ACC on Weizmann in Table 2) and consistent across feature extractors. The ablation study (Table 3) shows that each loss term contributes, and the sensitivity analysis covers the main hyper-parameters. The main reason for caution is that the theoretical inheritance from MCR2 is overclaimed: the objective actually optimized differs from the problem for which global optimality is proven, and the direct evidence of subspace alignment is a PCA plot of one sequence. These issues are fixable with additional analysis and experiments, so the significance is conditional.

major comments (3)
  1. [§3.2, Eqs. (9)–(13), Algorithm 1] The theoretical grounding of the UoS-alignment claim does not cover the objective actually optimized. The global optimality results cited for MCR2 ([52,58]) apply to max_Z ρ(Z)−ρc(Z|Π) with a fixed hard partition Π in Eq. (1), whereas TR2C minimizes −ρ(Z)+λ1ρc(Z|Γ)+λ2r(Z) in Eq. (9) with λ1≠1 and λ2>0, replaces Π by a doubly stochastic Γ obtained by Sinkhorn projection of Y~ᵀY~ in Eq. (12), and optimizes network parameters by gradient descent in Eq. (13). None of the cited theorems covers this composition, so the claim in §4.2 that TR2C 'explicitly ensures' that the learned representations align with a union of subspaces is currently an assertion rather than a consequence of the MCR2 principle. I request either a convergence or alignment analysis for the relaxed nonconvex problem, or an explicit de-emphasis of the theoretical guarantee together with the direct empirical test requested below.
  2. [§4.3, 'Quantitative evaluation of representations', Figures 2–3] The evidence that learned representations align with a UoS is qualitative or indirect. Figure 2 shows PCA projections of one sequence per dataset; Figure 3 shows that clustering algorithms perform better on the learned representations. Neither directly measures subspace structure, and since the loss explicitly minimizes the conditional coding rate, improved clustering of Z is partly by construction rather than independent evidence of UoS alignment. A direct test would be to fit a K-subspace model to Z, compute the normalized reconstruction residual per cluster or the principal angles between estimated per-class subspaces, and compare these quantities for TR2C, GCTSC, and raw HoG features across all sequences with standard deviations. Because the paper's central causal story is that the learned UoS structure drives the accuracy gain, this evidence is load-bearing and should be added.
  3. [§4.1, Appendix A.2, Figure 5] The evaluation protocol for λ1 and λ2 is not specified. Table A.1 lists per-dataset values and §4.1 states they are 'tuned independently for each dataset', but the paper does not state whether the tuning was performed on a validation split, by oracle access to test labels, or by a fixed grid, nor does it report variability of the final ACC/NMI over the tuning grid. Figure 5 shows sensitivity only on the first sequence of Weiz and UT, not the full benchmark. Given that the reported SOTA margins on some datasets are modest (e.g., 83.08 vs. 82.97 on MAD in Table 2, which is smaller than the reported standard deviation), the absence of a stated selection protocol makes it impossible to rule out test-set overfitting and should be corrected.
minor comments (5)
  1. [Algorithm 1] Line 4 contains the typo 'affinty' and missing spacing; it should read 'Compute affinity Γ(θ) by (12)'.
  2. [Figure 4] The caption says the thin lines use 'LSC', but the legend and text refer to 'LSR'; please make the acronym consistent.
  3. [Figure 5] The bottom-row x-axis is printed as values 0.001–0.13 under the label 's', while Eq. (7) defines s as an integer window size; if those panels show the coding precision ε, the axes should be relabeled.
  4. [Appendix A.1 / Table 1] The Weizmann entry in Table 1 lists 9 sequences while Appendix A.1 states the dataset contains 90 sequences; the paper should state explicitly that Table 1 reports the number of concatenated videos used after preprocessing.
  5. [§4.2, Table 2] Baseline numbers are cited from [9,63] without confidence intervals or a description of the evaluation protocol used to obtain them; adding this information would make the SOTA comparison easier to assess.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the HMS evaluation is external and self-contained; the UoS-alignment claim has a theorem-transfer gap, but that is a correctness risk, not a circular reduction.

full rationale

The paper's central empirical claim is grounded in clustering accuracy on five external benchmark datasets (Tables 2 and 5), with ground-truth labels used only for evaluation and never in the training loss. The derivation of Eq. (9) combines the MCR2 objective (Eq. (1), cited to [52,58]) with a temporal Laplacian regularizer and a collapse-preventing rate term; the cited optimality results apply to the hard-partition supervised problem, not to the relaxed neural/Sinkhorn objective in Eqs. (10)-(13). That is a support gap: the assertion that TR2C's learned representations align with a UoS is not established by the cited theorems for the actual objective. It is not a circular reduction, because the paper does not define UoS alignment as the value of Eq. (9), the affinity Gamma is learned rather than fixed to the ground-truth partition, and the reported ACC/NMI come from external labels. The qualitative PCA and downstream-clustering checks (Figs. 2-4) are empirical evidence about the learned representations rather than restatements of the loss. The ablation table is consistent with the loss components: the first row corresponds to Lrho + Lrhoc without the temporal term, which is essentially the MLC configuration cited as [11], so the table is not a renamed baseline. Self-citations such as [35] and [59] are design precedents and are not load-bearing proofs that force the conclusion. Hyper-parameter tuning per dataset is a reporting caveat rather than a fitted-input-called-prediction step. Overall, no step in the paper's derivation chain reduces by construction to its own inputs.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The paper's central claim depends on the UoS prior and the transfer of MCR2's alignment guarantee to a relaxed, network-parameterized objective, plus per-dataset selected weights λ1 and λ2 and known cluster counts. No invented entities are introduced.

free parameters (6)
  • λ1 (per-dataset loss weight, HMS HoG/VGG) = Weiz 0.1, Keck 0.1, UT 0.1, MAD 0.15, YouTube 1.0
    Tuned independently per dataset (Appendix A.2, Table A.1); no validation protocol stated.
  • λ2 (per-dataset loss weight, HMS HoG/VGG) = Weiz 12, Keck 10, UT 10, MAD 15, YouTube 2
    Tuned independently per dataset alongside λ1.
  • λ1 and λ2 for CLIP and TAS runs = CLIP: Weiz 0.1/12, Keck 0.1/10, YouTube 1/2; TAS: Breakfast 0.05/12, YouTube-Instr 0.05/20, 50Salads 0.05/15
    Separate per-dataset tuning on additional feature sets and temporal action segmentation datasets.
  • number of clusters K = 10 for Weiz/Keck/MAD/YouTube, 6 for UT; TAS varies
    Spectral clustering requires K; assumed known from dataset ground truth, standard for HMS benchmarks.
  • architecture dimensions dpre and d = d=64, dpre=512 for HMS; dpre varies for TAS
    Chosen architecture hyperparameters, fixed across datasets.
  • window size s, coding precision ε, learning rate η, iterations T = s=2, ε=0.1, η=5e-3, T=500 for HMS; T=100 for CLIP and TAS
    Fixed by inspection, reported in Appendix A.2.
assumptions (6)
  • domain assumption Frames of a video capturing consecutive human motions lie in a union of low-dimensional subspaces after an appropriate learned transformation.
    Section 3.1, Eqs. (5)-(9); the entire method is built on the UoS prior. Not proven for raw features; the paper argues raw HoG deviates from UoS, motivating learned Z.
  • standard math MCR2's rate reduction objective produces representations aligned with a union of orthogonal subspaces.
    Invoked via citations [52,58] for the supervised hard-partition problem; treated as a background theorem.
  • domain assumption Temporally neighboring frames are more likely to belong to the same motion, justifying the Laplacian regularizer with a fixed window.
    Section 3.1, Eqs. (6)-(7); borrowed from TSC [27].
  • ad hoc to paper The doubly stochastic relaxation Γ and Sinkhorn projection preserve the clustering properties of the original hard-partition problem.
    Section 3.2, Eqs. (10)-(12); no proof provided, the paper follows MLC [11].
  • ad hoc to paper Gradient descent over network parameters θ solves the nonconvex problem (9) well enough to reach useful solutions.
    Section 3.2, Algorithm 1; no convergence or landscape guarantees, standard empirical assumption.
  • domain assumption The number of motion segments K is known in advance.
    Section 3.2 and experimental setup; spectral clustering requires K.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Temporal Rate Reduction Clustering for Human Motion Segmentation." pith.science (2026). https://pith.science/paper/ROXIQZNT

@misc{pith2026250621249,
  author       = {Pith},
  title        = {Pith review of: Temporal Rate Reduction Clustering for Human Motion Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ROXIQZNT}},
  note         = {Machine review of arXiv:2506.21249}
}
abstract

Human Motion Segmentation (HMS), which aims to partition videos into non-overlapping human motions, has attracted increasing research attention recently. Existing approaches for HMS are mainly dominated by subspace clustering methods, which are grounded on the assumption that high-dimensional temporal data align with a Union-of-Subspaces (UoS) distribution. However, the frames in video capturing complex human motions with cluttered backgrounds may not align well with the UoS distribution. In this paper, we propose a novel approach for HMS, named Temporal Rate Reduction Clustering ($\text{TR}^2\text{C}$), which jointly learns structured representations and affinity to segment the sequences of frames in video. Specifically, the structured representations learned by $\text{TR}^2\text{C}$ enjoy temporally consistency and are aligned well with a UoS structure, which is favorable for addressing the HMS task. We conduct extensive experiments on five benchmark HMS datasets and achieve state-of-the-art performances with different feature extractors. The code is available at: https://github.com/mengxianghan123/TR2C.

Figures

Figures reproduced from arXiv: 2506.21249 by the authors.

Figure 1
Figure 1. The Framework of TR2C. Structured representations and affinity are jointly learned in TR2C to facilitate motion segmentation. press the representations which are drawn from a mixture of Gaussian distributions in lossy compression scenario [34]. To be specific, given a set of diagonal matrices Π = {Πj ∈ {0, 1} N×N } K j=1, where the m-th diagonal element of Πj indicates whether the m-th sample belongs to the j-th cla… view at source ↗
Figure 2
Figure 2. Visualization of features via PCA. First row: input HoG features. Second row: learned TR2C representations. Exper￾iments are conducted on the first sequence of each dataset. subspaces distribution of features is a key factor contribut￾ing to the state-of-the-art performance of TR2C. Quantitative evaluation of representations. To quantita￾tively evaluate the effectiveness of the learned representa￾tion of TR2C, we il… view at source ↗
Figure 3
Figure 3. Clustering accuracy of using HoG features, learned features, and learned affinity. Experiments are conducted on the first sequence of each dataset. Robustness evaluation of representations. Intuitively, if the representations align with a UoS structure, they will en￾joy strong robustness to the random noise corruption. To verify this, we corrupt the learned representations of TR2C, GCTSC [9] and HoG features by the … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Clustering accuracy of features under noise corrup￾tion. We test on features learned by TR2C, GCTSC, and HoG features, using EnSC (thick) and LSC (thin) for clustering. Ablation study. To study the effect of each component in the loss function, we conduct ablation stud…
Figure 5
Figure 5. Figure 5: Sensitivity to hyper-parameters. The sensitivity of TR2C with respect to λ1, λ2, s and ϵ is studied through experiments on the first sequence of the Weiz and UT dataset with three different random seeds [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 62 canonical work pages

  1. [1]

    A percep- tual prediction framework for self supervised event segmen- tation

    Sathyanarayanan N Aakur and Sudeep Sarkar. A percep- tual prediction framework for self supervised event segmen- tation. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 1197–1206,

    work page

  2. [2]

    Unsu- pervised learning from narrated instruction videos

    Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Josef Sivic, Ivan Laptev, and Simon Lacoste-Julien. Unsu- pervised learning from narrated instruction videos. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4575–4583, 2016. 8, 5

    work page 2016

  3. [3]

    Dual-side auto-encoder for high-dimensional time series seg- mentation

    Yue Bai, Lichen Wang, Yunyu Liu, Yu Yin, and Yun Fu. Dual-side auto-encoder for high-dimensional time series seg- mentation. In Proceedings of the IEEE International Confer- ence on Data Mining, pages 918–923, 2020. 2, 5, 6

    work page 2020

  4. [4]

    Human motion segmentation via velocity-sensitive dual-side auto-encoder

    Yue Bai, Lichen Wang, Yunyu Liu, Yu Yin, Hang Di, and Yun Fu. Human motion segmentation via velocity-sensitive dual-side auto-encoder. IEEE Transactions on Image Pro- cessing, 32:524–536, 2022. 2, 4, 5, 6

    work page 2022

  5. [5]

    Image clustering via the principle of rate reduction in the age of pretrained models

    Tianzhe Chu, Shengbang Tong, Tianjiao Ding, Xili Dai, Benjamin David Haeffele, Ren ´e Vidal, and Yi Ma. Image clustering via the principle of rate reduction in the age of pretrained models. In International Conference on Learning Representations, 2024. 2

    work page 2024

  6. [6]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in Neural Information Pro- cessing Systems, 26:2292–2300, 2013. 4

    work page 2013

  7. [7]

    Enhancing tempo- ral segmentation by nonlocal self-similarity

    Mariella Dimiccoli and Herwig Wendt. Enhancing tempo- ral segmentation by nonlocal self-similarity. In Proceedings of the IEEE International Conference on Image Processing , pages 3681–3685, 2019. 2

    work page 2019

  8. [8]

    Learning event rep- resentations for temporal segmentation of image sequences by dynamic graph embedding

    Mariella Dimiccoli and Herwig Wendt. Learning event rep- resentations for temporal segmentation of image sequences by dynamic graph embedding. IEEE Transactions on Image Processing, 30:1476–1486, 2020. 2, 5, 6

    work page 2020

Show all 64 references
  1. [9]

    Graph constrained data representation learning for human motion segmentation

    Mariella Dimiccoli, Llu ´ıs Garrido, Guillem Rodriguez- Corominas, and Herwig Wendt. Graph constrained data representation learning for human motion segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1460–1469, 2021. 2, 4, 5, 6, 7, 8

    2021

  2. [10]

    Tempo- ral action segmentation: An analysis of modern techniques

    Guodong Ding, Fadime Sener, and Angela Yao. Tempo- ral action segmentation: An analysis of modern techniques. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 46(2):1011–1030, 2023. 8

    2023

  3. [11]

    Haeffele

    Tianjiao Ding, Shengbang Tong, Kwan Ho Ryan Chan, Xili Dai, Yi Ma, and Benjamin D. Haeffele. Unsupervised manifold linearizing and clustering. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 5450–5461, 2023. 2, 4, 6

    2023

  4. [12]

    Fast and unsupervised action boundary detection for action seg- mentation

    Zexing Du, Xue Wang, Guoqing Zhou, and Qing Wang. Fast and unsupervised action boundary detection for action seg- mentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3323– 3332, 2022. 8

    2022

  5. [13]

    Sparse subspace cluster- ing

    Ehsan Elhamifar and Ren ´e Vidal. Sparse subspace cluster- ing. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 2790–2797,

  6. [14]

    Log- det heuristic for matrix rank minimization with applications to hankel and euclidean distance matrices

    Maryam Fazel, Haitham Hindi, and Stephen P Boyd. Log- det heuristic for matrix rank minimization with applications to hankel and euclidean distance matrices. InAmerican Con- trol Conference, pages 2156–2162, 2003. 4

    2003

  7. [15]

    The visual analysis of human movement: A survey

    Dariu M Gavrila. The visual analysis of human movement: A survey. Computer Vision and Image Understanding , 73 (1):82–98, 1999. 1

    1999

  8. [16]

    Probabilistic tem- poral subspace clustering

    Behnam Gholami and Vladimir Pavlovic. Probabilistic tem- poral subspace clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3066–3075, 2017. 2

    2017

  9. [17]

    Actions as space-time shapes.IEEE Trans- actions on Pattern Analysis and Machine Intelligence , 29 (12):2247–2253, 2007

    Lena Gorelick, Moshe Blank, Eli Shechtman, Michal Irani, and Ronen Basri. Actions as space-time shapes.IEEE Trans- actions on Pattern Analysis and Machine Intelligence , 29 (12):2247–2253, 2007. 4, 1

    2007

  10. [18]

    A cri- tique of self-expressive deep subspace clustering

    Benjamin D Haeffele, Chong You, and Ren ´e Vidal. A cri- tique of self-expressive deep subspace clustering. In Inter- national Conference on Learning Representations, 2021. 4

    2021

  11. [19]

    Sequential max-margin event detectors

    Dong Huang, Shitong Yao, Yi Wang, and Fernando De La Torre. Sequential max-margin event detectors. In European Conference on Computer Vision, pages 410–424,

  12. [20]

    Towards understanding action recognition

    Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3192–3199, 2013. 1

    2013

  13. [21]

    Recognizing hu- man actions by learning and matching shape-motion proto- type trees

    Zhuolin Jiang, Zhe Lin, and Larry Davis. Recognizing hu- man actions by learning and matching shape-motion proto- type trees. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 34(3):533–547, 2012. 1, 2, 4

    2012

  14. [22]

    The language of actions: Recovering the syntax and semantics of goal- directed human activities

    Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal- directed human activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 780–787, 2014. 8, 5

    2014

  15. [23]

    Unsupervised learning of action classes with contin- uous temporal embedding

    Anna Kukleva, Hilde Kuehne, Fadime Sener, and Jurgen Gall. Unsupervised learning of action classes with contin- uous temporal embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 12066–12074, 2019. 5

    2019

  16. [24]

    Learning realistic human actions from movies

    Ivan Laptev, Marcin Marszalek, Cordelia Schmid, and Ben- jamin Rozenfeld. Learning realistic human actions from movies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1–8, 2008. 5

    2008

  17. [25]

    Structured sparse subspace clustering: A joint affinity learning and sub- space clustering framework

    Chun-Guang Li, Chong You, and Ren ´e Vidal. Structured sparse subspace clustering: A joint affinity learning and sub- space clustering framework. IEEE Transactions on Image Processing, 26(6):2988–3001, 2017. 1

    2017

  18. [26]

    Deeper insights into graph convolutional networks for semi-supervised learn- ing

    Qimai Li, Zhichao Han, and Xiao-Ming Wu. Deeper insights into graph convolutional networks for semi-supervised learn- ing. In Proceedings of the AAAI conference on artificial in- telligence, 2018. 4

    2018

  19. [27]

    Temporal subspace clus- tering for human motion segmentation

    Sheng Li, Kang Li, and Yun Fu. Temporal subspace clus- tering for human motion segmentation. In Proceedings of 9 the IEEE/CVF International Conference on Computer Vi- sion, pages 4453–4461, 2015. 1, 2, 3, 4, 5, 6

    2015

  20. [28]

    Robust subspace clustering with indepen- dent and piecewise identically distributed noise modeling

    Yuanman Li, Jiantao Zhou, Xianwei Zheng, Jinyu Tian, and Yuan Yan Tang. Robust subspace clustering with indepen- dent and piecewise identically distributed noise modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8720–8729, 2019. 5, 6

    2019

  21. [29]

    Neural manifold clustering and embedding

    Zengyi Li, Yubei Chen, Yann LeCun, and Friedrich T Som- mer. Neural manifold clustering and embedding. arXiv preprint arXiv:2201.10000, 2022. 2

    2022 arXiv

  22. [30]

    Robust sub- space segmentation by low-rank representation

    Guangcan Liu, Zhouchen Lin, and Yong Yu. Robust sub- space segmentation by low-rank representation. In Inter- national Conference on Machine Learning , pages 663–670,

  23. [31]

    in the wild

    Jingen Liu, Jiebo Luo, and Mubarak Shah. Recognizing re- alistic actions from videos “in the wild”. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1996–2003, 2009. 4, 1

    1996

  24. [32]

    Self-supervised learning via maximum entropy coding

    Xin Liu, Zhongdao Wang, Ya-Li Li, and Shengjin Wang. Self-supervised learning via maximum entropy coding. Ad- vances in Neural Information Processing Systems , 35: 34091–34105, 2022. 4

    2022

  25. [33]

    Robust and efficient subspace segmentation via least squares regression

    Canyi Lu, Hai Min, Zhong-Qiu Zhao, Lin Zhu, De-Shuang Huang, and Shuicheng Yan. Robust and efficient subspace segmentation via least squares regression. In European Con- ference on Computer Vision, pages 347–360, 2012. 1, 2, 5, 6

    2012

  26. [34]

    Seg- mentation of multivariate mixed data via lossy coding and compression

    Yi Ma, Harm Derksen, Wei Hong, and John Wright. Seg- mentation of multivariate mixed data via lossy coding and compression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(9):1546–1562, 2007. 3, 4, 7

    2007

  27. [35]

    Exploring a principled framework for deep subspace clustering

    Xianghan Meng, Zhiyuan Huang, Wei He, Xianbiao Qi, Rong Xiao, and Chun-Guang Li. Exploring a principled framework for deep subspace clustering. In International Conference on Learning Representations, 2025. 4

    2025

  28. [36]

    A survey of advances in vision-based human motion capture and analysis

    Thomas B Moeslund, Adrian Hilton, and V olker Kr ¨uger. A survey of advances in vision-based human motion capture and analysis. Computer Vision and Image Understanding , 104(2-3):90–126, 2006. 1

    2006

  29. [37]

    Dynamic bayesian networks: repre- sentation, inference and learning

    Kevin Patrick Murphy. Dynamic bayesian networks: repre- sentation, inference and learning . University of California, Berkeley, 2002. 2

    2002

  30. [38]

    Maxime Oquab, Timoth ´ee Darcet, Th´eo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael ...

    2024

  31. [39]

    Vision-based human motion analysis: An overview

    Ronald Poppe. Vision-based human motion analysis: An overview. Computer Vision and Image Understanding, 108 (1-2):4–18, 2007. 1

    2007

  32. [40]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In International Conference on Machine Learning...

    2021

  33. [41]

    Spatio-temporal re- lationship match: Video structure comparison for recogni- tion of complex human activities

    Michael S Ryoo and Jake K Aggarwal. Spatio-temporal re- lationship match: Video structure comparison for recogni- tion of complex human activities. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 1593–1600, 2009. 1, 2, 4

    2009

  34. [42]

    Temporally-weighted hierarchical clustering for unsupervised action segmentation

    Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11225–1123...

    2021

  35. [43]

    Normalized cuts and image segmentation

    Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888–905, 2000. 4, 6

    2000

  36. [44]

    Very deep convo- lutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convo- lutional networks for large-scale image recognition. In In- ternational Conference on Learning Representations , 2015. 5

    2015

  37. [45]

    Probabilistic model-based clustering of multivariate and sequential data

    Padhraic Smyth. Probabilistic model-based clustering of multivariate and sequential data. In Proceedings of the In- ternational Workshop on AI and Statistics , pages 299–304,

  38. [46]

    Hierarchical vector quantization for unsuper- vised action segmentation

    Federico Spurio, Emad Bahrami, Gianpiero Francesca, and Juergen Gall. Hierarchical vector quantization for unsuper- vised action segmentation. In Proceedings of the AAAI Con- ference on Artificial Intelligence, pages 6996–7005, 2025. 8, 6

    2025

  39. [47]

    Sebastian Stein and Stephen J. McKenna. Combining em- bedded accelerometers with computer vision for recognizing food preparation activities. In Proceedings of the ACM In- ternational Joint Conference on Pervasive and Ubiquitous Computing, pages 729–738, 2013. 8, 5

    2013

  40. [48]

    Subspace cluster- ing for sequential data

    Stephen Tierney, Junbin Gao, and Yi Guo. Subspace cluster- ing for sequential data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1019–1026, 2014. 1, 2, 3, 5, 6

    2014

  41. [49]

    Action recognition with improved trajectories

    Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision , pages 3551– 3558, 2013. 5

    2013

  42. [50]

    Learning trans- ferable subspace for human motion segmentation

    Lichen Wang, Zhengming Ding, and Yun Fu. Learning trans- ferable subspace for human motion segmentation. In Pro- ceedings of the AAAI Conference on Artificial Intelligence ,

  43. [51]

    Low-rank transfer human motion segmentation

    Lichen Wang, Zhengming Ding, and Yun Fu. Low-rank transfer human motion segmentation. IEEE Transactions on Image Processing, 28(2):1023–1034, 2018. 1, 2, 4, 5, 6

    2018

  44. [52]

    A global geometric analysis of maximal coding rate reduction

    Peng Wang, Huikang Liu, Druv Pai, Yaodong Yu, Zhihui Zhu, Qing Qu, and Yi Ma. A global geometric analysis of maximal coding rate reduction. In International Conference on Machine Learning, 2024. 2, 3

    2024

  45. [53]

    Support structure representation learning for sequential data cluster- ing

    Xiumei Wang, Dingning Guo, and Peitao Cheng. Support structure representation learning for sequential data cluster- ing. Pattern Recognition, 122:108326, 2022. 2 10

    2022

  46. [54]

    Coseg: Cognitively inspired unsupervised generic event segmenta- tion

    Xiao Wang, Jingen Liu, Tao Mei, and Jiebo Luo. Coseg: Cognitively inspired unsupervised generic event segmenta- tion. IEEE Transactions on Neural Networks and Learning Systems, 35(9):12507–12517, 2023. 8

    2023

  47. [55]

    Mixtures of arma models for model-based time series clustering

    Yimin Xiong and Dit-Yan Yeung. Mixtures of arma models for model-based time series clustering. In Proceedings of the IEEE International Conference on Data Mining , pages 717–720, 2002. 2

    2002

  48. [56]

    Temporally consistent unbal- anced optimal transport for unsupervised action segmenta- tion

    Ming Xu and Stephen Gould. Temporally consistent unbal- anced optimal transport for unsupervised action segmenta- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 14618–14627,

  49. [57]

    Oracle based active set algorithm for scalable elastic net subspace clustering

    Chong You, Chun-Guang Li, Daniel Robinson, and Ren´e Vi- dal. Oracle based active set algorithm for scalable elastic net subspace clustering. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 3928–3937, 2016. 1, 6

    2016

  50. [58]

    Learning diverse and discriminative rep- resentations via the principle of maximal coding rate reduc- tion

    Yaodong Yu, Kwan Ho Ryan Chan, Chong You, Chaobing Song, and Yi Ma. Learning diverse and discriminative rep- resentations via the principle of maximal coding rate reduc- tion. Advances in Neural Information Processing Systems , 33:9422–9434, 2020. 1, 2, 3, 4

    2020

  51. [59]

    Learning a self-expressive network for subspace cluster- ing

    Shangzhi Zhang, Chong You, Ren ´e Vidal, and Chun-Guang Li. Learning a self-expressive network for subspace cluster- ing. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 12393–12403,

  52. [60]

    Un- supervised discovery of facial events

    Feng Zhou, Fernando De la Torre, and Jeffrey F Cohn. Un- supervised discovery of facial events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2574–2581, 2010. 2

    2010

  53. [61]

    Hierarchical aligned cluster analysis for temporal clustering of human motion

    Feng Zhou, Fernando De la Torre, and Jessica K Hodgins. Hierarchical aligned cluster analysis for temporal clustering of human motion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(3):582–596, 2012. 2

    2012

  54. [62]

    Multi-mutual consistency induced transfer subspace learning for human motion segmentation

    Tao Zhou, Huazhu Fu, Chen Gong, Jianbing Shen, Ling Shao, and Fatih Porikli. Multi-mutual consistency induced transfer subspace learning for human motion segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10277–10286, 2020...

    2020

  55. [63]

    Consistency and diversity induced human motion segmentation

    Tao Zhou, Huazhu Fu, Chen Gong, Ling Shao, Fatih Porikli, Haibin Ling, and Jianbing Shen. Consistency and diversity induced human motion segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence , 45(1):197–210,

  56. [64]

    compressed

    Qiang Zhu, Mei-Chen Yeh, Kwang-Ting Cheng, and Shai Avidan. Fast human detection using a cascade of histograms of oriented gradients. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 1491–1498, 2006. 5 11 Temporal Rate Reduction Clu...

    2006

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.