Pith. sign in

REVIEW 2 major objections 4 minor 26 references

Beyond Pairwise Relations: Dynamic Manipulation Hypergraphs for Vision-Based Human Activity Recognition

T0 review · 2 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Manipulation actions are better recognized when the entities that jointly participate—hands, objects, tools, and surfaces—are bound into single time-varying hyperedges rather than pairwise links.

desk verdict A thoughtful dynamic-hypergraph paper with honest controlled experiments, but the 'matched' pairwise baseline is underspecified enough that the headline gains aren't fully accountable yet. read the letter →

arxiv 2607.14350 v2 pith:DW3W7YTY submitted 2026-07-15 cs.CV

Fatemeh Ziaeetabar This is my paper
classification cs.CV
keywords humanactivityrecognitiondynamichypergraphhigher-orderrelationsegocentricvisionmanipulationunderstandinghuman-objectinteractionactionneuralnetwork
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

Fine-grained manipulation actions like cutting or placing are defined by how several entities—hands, objects, tools, surfaces—come together, not by isolated pairs. This paper argues that representing each such configuration as a single time-varying hyperedge, rather than decomposing it into pairwise relations, improves recognition. The authors build a dynamic manipulation hypergraph: at each temporal step, up to five entity roles are detected, candidate hyperedges from four templates are ranked by contact, proximity, and motion coupling, and a message-passing network reasons over them before temporal attention pools the result. On two benchmarks, the model raises higher-order action F1 by 6.9 and 9.5 points over a matched pairwise graph, and by 4.4 and 5.8 over a static hypergraph; top-ranked hyperedges also align with contact-rich intervals in qualitative analysis. If the claim holds, higher-order relational structure is a meaningful and underused signal for fine-grained activity recognition.

What carries the argument

The dynamic manipulation hypergraph. Nodes are manipulation-relevant entities drawn from five canonical roles (left hand, right hand, object, tool, supporting surface); hyperedges are candidate multi-entity configurations instantiated from four fixed templates (hand–object–surface, left-hand–right-hand–object, active-hand–tool–object–surface) and ranked by scores built from proximity, contact, and motion-coupling predicates. The network alternates node-to-hyperedge and hyperedge-to-node message passing with relevance-biased attention, then aggregates steps via temporal attention; the importance score is the product of temporal attention and within-step hyperedge relevance.

What would settle it

If a fully automatic entity-extraction pipeline (no annotation-assisted masks) is used on a third egocentric manipulation benchmark, and the dynamic hypergraph no longer beats the matched pairwise graph on HO-F1, then the reported gains depend on the annotation protocol rather than on higher-order structure.

Watch

Extended reading notes

Core claim

The paper's central claim is that manipulation activities are best recognized by representing them as sequences of hypergraphs, where each hyperedge binds all entities jointly involved in a configuration (e.g., active hand–tool–object–surface) at a given moment, and where these hyperedges are re-selected over time based on proximity, contact, and motion-coupling evidence. This dynamic higher-order representation, after node-to-hyperedge and hyperedge-to-node message passing and temporal attention, produces a video-level feature for classification. Compared with a matched pairwise graph using identical entity inputs and reasoning depth, the proposed method improves HO-F1 by 6.9 percentage poi

Load-bearing premise

The framework's benefit rests on the assumption that the five entity roles can be reliably detected and assigned from video without using the activity label; if an entity is missing, the hypergraph falls back to pooled node features and the advantage disappears.

Editorial extensions

If this is right

  • If the central claim is right, recognition systems for manipulation-heavy video should represent coordinated entity groups as higher-order units rather than only pairwise links; the reported 6.9/9.5 HO-F1 gains on the two benchmarks quantify the headroom from this change alone.
  • Time-varying hypergraph construction matters beyond higher-order structure alone, since the static hypergraph underperforms the dynamic one by 4.4/5.8 points; per-step re-selection of the active configuration carries independent signal.
  • The largest family-wise gains are for tool-mediated and bimanual actions (up to 9.5 points), suggesting these actions are precisely where pairwise decomposition loses the jointly-participating configuration.
  • The class-agnostic hyperedge-importance scores provide a non-causal but temporally aligned map of which multi-entity configurations the model relies on, opening a route to relational explanations and failure analysis (e.g., missing-surface cases) without claiming causality.

Reading between the lines

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

  • The paper leaves open whether the four fixed templates are necessary; a testable extension is to learn hyperedge candidates from data and compare coverage on held-out action classes.
  • The same dynamic-hypergraph construction could transfer to group activity or human-robot collaboration, where the jointly acting unit is also larger than a pair; the intended lesson is about finding the right participant group per timestep, not hypergraphs per se.
  • Because the reported gains are measured with annotation-assisted entity localization, they are an upper bound for fully automatic pipelines; re-measuring under automatic detection errors would quantify the real-world headroom.
  • The class-agnostic importance scores, though not causal, could be repurposed as weak supervision for localizing manipulation-relevant moments or for selecting informative clips in a video-language model; the paper does not explore this.
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

2 major / 4 minor

Summary. This paper proposes a dynamic manipulation hypergraph framework for fine-grained activity recognition. The model extracts up to five entity roles (left/right hand, manipulated object, tool, supporting surface), encodes appearance, spatial, motion, and semantic-role features, instantiates candidate hyperedges from four fixed templates using contact, proximity, and motion-coupling predicates, ranks them with a learned relevance score, keeps the top K per temporal step, and applies hypergraph message passing followed by temporal attention. It also outputs class-agnostic hyperedge-importance scores. Quantitative experiments on EPIC-KITCHENS-100/VISOR and Assembly101 report HO-F1 improvements over a matched pairwise graph (6.9 and 9.5 points) and over a static hypergraph (4.4 and 5.8 points), with ARCTIC used for qualitative contact analysis.

Significance. If the controlled comparisons are indeed matched, the paper makes a credible contribution: it provides evidence that time-varying higher-order relational structure helps fine-grained manipulation recognition beyond pairwise graphs, and it does so with a clear formulation and consistent ablations. Strengths include the use of identical entity inputs across the relational baselines, explicit caveats that hyperedge-importance scores are attention-based rather than causal, and acknowledgment of the dependence on reliable entity localization. The main significance, however, hinges on the exact architecture of the matched pairwise and static hypergraph baselines and on the transparency of the custom EPIC split and the HO-F1 class set; without these, the magnitude of the reported gains cannot be independently assessed.

major comments (2)
  1. [§4.3, Table 3, Eq. (23), Alg. 1 lines 26-30, Eq. (34)] The central claim rests on the 'matched pairwise graph' baseline, but the manuscript does not specify whether this baseline includes the same per-template relevance scoring (Eq. 23), learnable template embeddings mu_tau(j), top-K selection (Alg. 1 lines 26-30), and relevance-biased attention lambda*alpha (Eq. 34). If any of these are omitted, the reported HO-F1 gains conflate higher-order structure with additional model capacity or selectivity. The static hypergraph baseline is similarly underspecified: 'template identities are selected once' does not state how many templates are kept or whether the K=3 per-step selection is matched. The ablations in Table 5 never add equivalent components to the baselines, so they cannot isolate 'higher-orderness.' Please provide exact baseline architectures and, ideally, code and released splits.
  2. [§4.1, Eq. (49), Table 3] The headline metric HO-F1 is computed only on a predefined subset of classes, but the actual class list is not given. The EPIC-KITCHENS-100/VISOR experiment uses a custom internal split (8,920 clips) that is not released. Because the HO-F1 class set is selected by the authors and the split is non-standard, the main results cannot be independently reproduced or checked for selection bias. Please release the full class list, the split-generation procedure, and the splits/code.
minor comments (4)
  1. [Eq. (17)] There is a formatting artifact in the norm expression of the motion-coupling score; please fix the typesetting.
  2. [§3.2, Eq. (20)] The set V_hand^t is used but not defined; define it as the subset of node indices labeled as hands.
  3. [§4.2] The hyperparameters tau_d, tau_c, lambda, K, and L are fixed values; a brief sensitivity study or at least the tuning range would strengthen the claim that the results are not sensitive to these choices.
  4. [§4.7] The ARCTIC analysis is informal and uses only the top-ranked hyperedge per clip (R=1). Consider reporting a simple quantitative alignment measure (e.g., overlap between high-importance frames and annotated contact intervals) or describing the analysis as purely illustrative.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: held-out evaluation and independent baselines support the empirical claims.

full rationale

The paper's central claim is an empirical accuracy comparison on held-out test labels (EPIC-KITCHENS-100/VISOR and Assembly101) under a fixed protocol. Hypergraph construction uses label-independent role assignment and predicate scores; the classification head is trained with cross-entropy on ground-truth labels, so the recognition outcome is not defined by the hyperedge-importance scores. HO-F1 is computed on a predefined subset of classes selected by activity definition rather than by model performance, so this is an evaluation-scope choice, not a circular derivation. The 'matched pairwise graph' and 'static hypergraph' baselines are controlled comparisons; although their implementation details (per-template scoring, top-K selection, relevance-biased attention) are underspecified, underspecification is an internal-validity/correctness risk, not a circularity, because the paper does not define the baseline in terms of the proposed model's outputs. Prior self-citations [5,6,23,24] are background and motivation, not load-bearing justifications of the empirical gains. The ARCTIC analysis is explicitly qualitative and non-causal, so attention-based importance scores are not presented as predictions of ground truth. No step in the derivation sets an input equal to the claimed output; therefore no circularity is established.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method rests on a fixed template set and reliable entity roles; these are domain assumptions rather than derived facts. The free parameters are hyperparameters selected on validation; they are not fitted to test labels but could affect the relative gain of the dynamic hypergraph.

free parameters (6)
  • τ_d (proximity scale) = 0.20
    Hand-set/validation-tuned; controls exponential decay of proximity score in Eq. (14).
  • τ_c (contact-distance scale) = 0.01
    Hand-set/validation-tuned; controls exponential decay of contact score in Eq. (16).
  • λ (relevance coefficient) = 1
    Fixed by hand; balances hyperedge relevance score α in attention weights, Eq. (34).
  • K (max hyperedges per temporal step) = 3
    Hand-set; limits retained hyperedges per step (§4.2).
  • L (reasoning layers) = 3
    Hand-set; depth of node-to-hyperedge and hyperedge-to-node message passing (§4.2).
  • R (reported hyperedges) = 1
    For qualitative visualization only (§4.2).
assumptions (5)
  • domain assumption Four manipulation templates (τ1–τ4 in Eq. 18) are sufficient to instantiate meaningful higher-order configurations for the evaluated activity classes.
    If an action's multi-entity configuration falls outside these templates, no hyperedge is formed (Section 3.3); the method's advantage relies on template coverage.
  • domain assumption Hands, manipulated object, tool, and supporting surface are the canonical manipulation entities, and role assignment can be performed without using activity labels.
    Section 3.2 assumes label-independent role selection; errors or missing roles degrade the representation (acknowledged in Section 5).
  • domain assumption The predefined higher-order class subset (surface-dependent, tool-mediated, bimanual, contact/state-change) is the appropriate target metric for evaluating higher-order relational modeling.
    HO-F1 is defined only over this subset (Eq. 49); the reported headline gains are concentrated there.
  • domain assumption The annotation-assisted entity-localization protocol on EPIC-KITCHENS-100/VISOR (using VISOR masks) is a valid testbed for relational reasoning, and the controlled baselines provide fair comparison.
    Using ground-truth masks for hands/objects may favor relational methods; the author argues identical inputs for all relational baselines mitigate this.
  • domain assumption Frozen ImageNet-pretrained ResNet-50 features are sufficient entity representations for manipulation recognition.
    The appearance backbone is frozen; the method's gains stem from relational structure, not from stronger visual features.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Pairwise Relations: Dynamic Manipulation Hypergraphs for Vision-Based Human Activity Recognition." pith.science (2026). https://pith.science/paper/DW3W7YTY

@misc{pith2026260714350,
  author       = {Pith},
  title        = {Pith review of: Beyond Pairwise Relations: Dynamic Manipulation Hypergraphs for Vision-Based Human Activity Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DW3W7YTY}},
  note         = {Machine review of arXiv:2607.14350}
}
read the original abstract

Fine-grained manipulation recognition requires modeling evolving relations among hands, objects, tools, and supporting surfaces. Conventional graph-based methods use pairwise edges that can fragment a coordinated event into disconnected binary relations. We propose a dynamic manipulation hypergraph framework that represents multi-entity configurations as higher-order relational units. At each temporal step, relevant entities are encoded using appearance, spatial, motion, and semantic-role features. Hyperedge candidates are instantiated and ranked using proximity, contact, and motion-coupling predicates. A hypergraph reasoning network performs node-to-hyperedge and hyperedge-to-node message passing, followed by temporal attention over the evolving interaction structure. The framework provides class-agnostic hyperedge-importance scores that identify entity configurations and temporal intervals emphasized by the model without treating them as causal explanations. Quantitative evaluation is conducted on EPIC-KITCHENS-100/VISOR and Assembly101 under an annotation-assisted entity-localization protocol. Video-only and entity-based methods provide contextual comparisons, while a matched pairwise graph and a static hypergraph serve as the principal controlled baselines because they use identical entity inputs and comparable relational settings. The proposed method improves HO-F1 over the matched pairwise graph by 6.9 percentage points on EPIC-KITCHENS-100/VISOR and 9.5 points on Assembly101, and exceeds the static hypergraph by 4.4 and 5.8 points, respectively. Qualitative analysis on ARCTIC further shows correspondence between highly ranked hyperedges and contact-rich manipulation intervals. These results demonstrate the value of time-varying higher-order relational modeling for fine-grained manipulation activity recognition.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 1 canonical work pages

  1. [1]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Wang, L., Huang, B., Zhao, Z., Tong, Z., He, Y., Wang, Y., Wang, Y., Qiao, Y.: Videomae v2: Scaling video masked autoencoders with dual masking. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14549–14560 (2023)

    2023

  2. [2]

    In: Computer Vision – ECCV 2024, pp

    Wang, Y., Li, K., Li, X., Yu, J., He, Y., Wang, C., Chen, G., Pei, B., Zheng, R., Wang, Z., Shi, Y., Jiang, T., Li, T., Xu, J., Zhang, H., Huang, Y., Qiao, Y., Wang, Y., Wang, L.: Internvideo2: Scaling foundation models for multimodal video understanding. In: Computer Vision – ECCV 2024, pp. 396–416 (2024)

    2024

  3. [3]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021)

    Kwon, T., Tekin, B., St¨ uhmer, J., Bogo, F., Pollefeys, M.: H2o: Two hands manip- ulating objects for first person interaction recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021)

    2021

  4. [4]

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

    Rodin, I., Furnari, A., Min, K., Tripathi, S., Farinella, G.M.: Action scene graphs for long-form understanding of egocentric videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 18622–18632 (2024)

    2024

  5. [5]

    IEEE Access12, 180328–180360 (2024)

    Ziaeetabar, F., Tamosiunaite, M., W¨ org¨ otter, F.: A hierarchical graph-based approach for recognition and description generation of bimanual actions in videos. IEEE Access12, 180328–180360 (2024)

    2024

  6. [6]

    IEEE Access13, 201990–202009 (2025)

    Ziaeetabar, F., W¨ org¨ otter, F.: Adaptive multimodal graph reasoning with founda- tion models for fine-grained action recognition. IEEE Access13, 201990–202009 (2025)

    2025

  7. [7]

    International Journal of Computer Vision130(1), 33–55 (2022) 30

    Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Rescaling egocentric vision: Col- lection, pipeline and challenges for EPIC-KITCHENS-100. International Journal of Computer Vision130(1), 33–55 (2022) 30

    2022

  8. [8]

    In: Advances in Neural Information Processing Systems, vol

    Darkhalil, A., Shan, D., Zhu, B., Ma, J., Kar, A., Higgins, R., Fidler, S., Fouhey, D., Damen, D.: EPIC-KITCHENS VISOR Benchmark: VIdeo Segmentations and Object Relations. In: Advances in Neural Information Processing Systems, vol. 35 (2022). Datasets and Benchmarks Track

    2022

Show all 26 references
  1. [9]

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

    Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 21096–21106 (2022)

    2022

  2. [10]

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

    Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: Arctic: A dataset for dexterous bimanual hand-object manipulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 12943–12954 (2023)

    2023

  3. [11]

    IEEE Transactions on Pattern Analysis and Machine Intelligence45(6), 6703–6714 (2023) https://doi.org/10.1109/TPAMI.2021.3055233

    Dessalene, E., Devaraj, C., Maynord, M., Fermuller, C., Aloimonos, Y.: Fore- casting action through contact representations from first person video. IEEE Transactions on Pattern Analysis and Machine Intelligence45(6), 6703–6714 (2023) https://doi.org/10.1109/TPAMI.2021.3055233

    2023

  4. [12]

    In: Computer Vision – ECCV 2022, pp

    Zhang, L., Zhou, S., Stent, S., Shi, J.: Fine-grained egocentric hand-object seg- mentation: Dataset, model, and applications. In: Computer Vision – ECCV 2022, pp. 127–145 (2022). https://doi.org/10.1007/978-3-031-19818-2 8

    2022 doi

  5. [13]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (2019)

    Feng, Y., You, H., Zhang, Z., Ji, R., Gao, Y.: Hypergraph neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence (2019)

    2019

  6. [14]

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

    Perrett, T., Darkhalil, A., Sinha, S., Emara, O., Pollard, S., Parida, K.K., Liu, K., Gatti, P., Bansal, S., Flanagan, K., Chalk, J., Zhu, Z., Guerrier, R., Abdelazim, F., Zhu, B., Moltisanti, D., Wray, M., Doughty, H., Damen, D.: Hd-epic: A highly- detailed egocentric video d...

    2025

  7. [15]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (2018)

    Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence (2018)

    2018

  8. [16]

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

    Wang, L., Koniusz, P.: 3mformer: Multi-order multi-mode transformer for skeletal action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5620–5631 (2023)

    2023

  9. [17]

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

    Zhou, Y., Yan, X., Cheng, Z.-Q., Yan, Y., Dai, Q., Hua, X.-S.: Blockgcn: Redefine topology awareness for skeleton-based action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2049– 2058 (2024)

  10. [18]

    arXiv preprint arXiv:2211.09590 (2022)

    Zhou, Y., Cheng, Z.-Q., Li, C., Fang, Y., Geng, Y., Xie, X., Keuper, M.: 31 Hypergraph transformer for skeleton-based action recognition. arXiv preprint arXiv:2211.09590 (2022)

    2022 arXiv

  11. [19]

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

    Ray, A., Raj, A., Kolekar, M.H.: Autoregressive adaptive hypergraph transformer for skeleton-based activity recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 9690–9699 (2025)

    2025

  12. [20]

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

    Nguyen, T.-T., Nguyen, P., Cothren, J., Yilmaz, A., Luu, K.: Hyperglm: Hyper- graph for video scene graph generation and anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 29150–29160 (2025)

    2025

  13. [21]

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

    Liu, Y., Liu, Y., Jiang, C., Lyu, K., Wan, W., Shen, H., Liang, B., Fu, Z., Wang, H., Yi, L.: Hoi4d: A 4d egocentric dataset for category-level human-object inter- action. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 21013–21022 (2022)

    2022

  14. [22]

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

    Shiota, T., Takagi, M., Kumagai, K., Seshimo, H., Aono, Y.: Egocentric action recognition by capturing hand-object contact and object state. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 6527–6537 (2024)

    2024

  15. [23]

    Machine Vision and Applications35(4), 64 (2024)

    Ziaeetabar, F., Safabakhsh, R., Momtazi, S., Tamosiunaite, M., W¨ org¨ otter, F.: Multi sentence description of complex manipulation action videos. Machine Vision and Applications35(4), 64 (2024)

    2024

  16. [24]

    arXiv preprint arXiv:2508.01465 (2025) https://doi.org/10.48550/arXiv.2508.01465 arXiv:2508.01465 [cs.CV]

    Ziaeetabar, F.: EfficientGFormer: Multimodal brain tumor segmentation via pruned graph-augmented transformer. arXiv preprint arXiv:2508.01465 (2025) https://doi.org/10.48550/arXiv.2508.01465 arXiv:2508.01465 [cs.CV]

    2025 doi arXiv

  17. [25]

    In: Computer Vision – ECCV 2024 (2024)

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., Zhang, L.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: Computer Vision – ECCV 2024 (2024)

    2024

  18. [26]

    In: International Conference on Learning Representations (2025) 32

    Ravi, N., Gabeur, V., Hu, Y.-T., Hu, R., Ryali, C., Ma, T., Khedr, H., R¨ adle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.-Y., Girshick, R., Doll´ ar, P., Feichtenhofer, C.: SAM 2: Segment anything in images and videos. In: Internatio...

    2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.