Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Bounding boxes and motion beat heavy vision models on JRDB

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 ↗

On JRDB, geometry plus optical flow alone reaches 84.3% accuracy for classifying walking/standing/sitting pairs, outperforming the same pipeline augmented with a frozen visual backbone, and transfers zero-shot to a lawnmower platform.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection The core claim—geometry+motion beats frozen appearance on JRDB—is supported; but missing Stage 1 recall and end-to-end metrics mean the headline accuracy may not reflect deployed performance. the 4 major comments →

arxiv 2602.22346 v2 pith:F2GG4S5C submitted 2026-02-25 cs.RO

A Pairwise Human-Human Interaction Detection and Recognition Framework for Mobile Service Robots

Mengyu Liang , Iolanda Leite , Sarah Gillet This is my paper
classification cs.RO MSC 68T4068T45
keywords pairwise interaction recognitionmobile service robotssocial navigationbounding box geometryoptical flowrelation networkJRDB datasetzero-shot transfer
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.

The reading

This paper argues that for mobile service robots, understanding human-human interaction is primarily a perception problem under limited sensing and compute, not a fine-grained recognition problem. It proposes that pairwise interactions are the minimal sufficient perceptual unit, and introduces a two-stage pipeline: first filter candidate interacting pairs using bounding-box geometry, then classify coarse interaction types using optical-flow motion cues and a relation network. On the JRDB benchmark, geometry plus motion alone reaches 84.3% accuracy at just 1.12 Mb FLOPs, outperforming every configuration that adds a frozen EfficientNet backbone, whose best accuracy is 80.0% with 5.80 Gb FLOPs. The paper concludes that simple geometric and motion cues dominate coarse-grained interaction recognition, making expensive appearance models unnecessary for robot-centric social understanding.

Core claim

The central claim is that simple geometric and motion cues provide a practical and efficient basis for interaction-aware perception in mobile service robots. Concretely, on JRDB the geometry+motion configuration achieves 84.3% accuracy with 1.12 Mb FLOPs, higher than any configuration using a frozen EfficientNet appearance backbone, and motion features alone yield 78.4% accuracy while geometry alone gives 59.3%. Adding appearance features does not improve accuracy and often degrades it, indicating redundancy with motion and geometry. The paper also shows zero-shot transfer to a lawnmower-collected dataset with 96.5% precision in interaction detection and real-time processing at 44 fps, suppo

What carries the argument

The two-stage framework. Stage 1 extracts a 7D geometric feature vector from person bounding boxes (horizontal gap, height ratio, ground distance, vertical overlap, area ratio, center distance, vertical gap), applies a learnable reweighting transformation, and passes it through a shallow MLP to propose interacting pairs while prioritizing recall. Stage 2 computes a 10D geometric-motion feature from bounding boxes and Farneback optical flow, including motion magnitude statistics, dominant direction, and an interaction synchrony feature that aggregates four similarity measures; optionally fused with frozen EfficientNet appearance features, these go through a relation network that combines per-

Load-bearing premise

Stage 1 must recall nearly all genuinely interacting pairs, so that filtering by bounding-box geometry does not discard interacting pairs before classification; the paper reports precision but not recall for this stage.

What would settle it

Measure Stage 1 recall and end-to-end detection+classification accuracy on JRDB using the released code. If Stage 1 recall is substantially below the reported Stage 2 accuracy, the deployed system underperforms the headline figures. Alternatively, find an interaction type (e.g., two people sitting close together with no relative motion) that is geometrically indistinguishable from two non-interacting people nearby; if such pairs are systematically misclassified in practice, the sufficiency claim is bounded.

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

If this is right

  • Mobile service robots can run interaction perception on-board in real time with negligible compute, enabling socially aware navigation without heavy visual models.
  • Skeleton-free, appearance-free interaction recognition is feasible outdoors under occlusion, low resolution, and motion blur, which are common on mobile platforms.
  • Pairwise interaction predictions can be aggregated into group structures, matching or beating unsupervised clustering baselines on CAD and approaching fully supervised group methods at lower complexity.
  • Dropping appearance features entirely simplifies the perception stack and reduces model size with no accuracy penalty on this coarse 3-class task.
  • The coarse interaction categories (walking, standing, sitting together) appear to capture the majority of JRDB's label structure, since other labels co-occur with these dominant classes.

Where Pith is reading between the lines

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

  • The practical deployment risk is Stage 1 recall: the paper reports only precision (96.5%) for zero-shot detection and no recall or end-to-end detection+classification on JRDB; if many true interacting pairs are filtered out early, real-world accuracy will fall below the 84.3% classification figure.
  • The sufficiency of geometry and motion may be specific to coarse, kinematically defined interactions; fine-grained types (e.g., conversing vs. waiting in place) could still require appearance cues, so the claim should not be over-generalized.
  • The same lightweight approach could likely transfer to other mobile platforms (delivery robots, drones) with minimal adjustment, since bounding boxes and optical flow are platform-agnostic; a quick test would be training on JRDB and evaluating on another outdoor robot dataset.
  • Motion compensation for egomotion appears to be a key fragility: the lawnmower's own motion caused sitting-together misclassifications, suggesting that explicit ego-motion subtraction or longer temporal windows would be a natural next step.
  • The finding that appearance features are redundant is itself testable: a controlled experiment with visually distinct but kinematically identical interactions (e.g., two people standing close vs. a mannequin and a person) would directly probe the limits of the sufficiency claim.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper proposes a two-stage pipeline for pairwise human-human interaction detection and classification aimed at mobile service robots. Stage 1 uses seven-dimensional bounding-box geometry features with a learnable reweighting and a small MLP to propose candidate interacting pairs. Stage 2 classifies those pairs into walking together, standing together, or sitting together using frozen EfficientNet appearance features together with optical-flow-derived motion and geometric features through a relation network. Experiments on JRDB show 84.3% accuracy for geometry+motion without appearance features at 1.12 Mb FLOPs, on CAD the method is compared against group-activity baselines, and a zero-shot deployment on a lawnmower-collected dataset is reported. The authors conclude that simple geometric and motion cues are a practical and efficient basis for interaction-aware perception.

Significance. If the results hold, this is a useful existence proof that lightweight, non-skeletal, non-appearance features can support coarse interaction recognition for robots, with a favorable accuracy-efficiency trade-off. The paper's strengths include a concrete two-stage design, an ablation study with multiple seeds for the main configurations, evaluation on a public benchmark plus a real platform, and released code. However, the significance is tempered by missing metrics: no Stage 1 recall and no end-to-end detection+classification evaluation are reported, so the deployed performance claim is not yet established.

major comments (4)
  1. [III-A1, IV-G, Table I] The headline 84.3% accuracy is a Stage 2 classification result on pairs that survived Stage 1. The paper explicitly states in Section III-A1 that Stage 1 is designed to 'prioritize recall over precision,' yet no Stage 1 recall is reported on JRDB, and Section IV-G reports only a precision of 96.5% for the zero-shot lawnmower deployment. Precision does not bound recall. If Stage 1 filters out genuine interacting pairs, the end-to-end detection+classification performance is lower than Table I suggests. This is load-bearing for the central claim that the framework is practical for mobile robots, where missing an interaction is the safety-relevant error. Please report Stage 1 recall on JRDB and an end-to-end metric (e.g., accuracy/F1 evaluated over all person pairs, counting Stage 1 misses as errors) on JRDB, and at minimum also report recall on the lawnmower set.
  2. [IV-A, IV-E, Table I] The JRDB train/validation/test split is never specified. The readers cannot tell whether the reported numbers are on a held-out test set or a validation set, nor how many videos/frames are used. In addition, several rows of Table I (mobilenet, resnet18, vgg16, alexnet) report no error bars, while the main rows show ± values over three seeds. Without a defined split and error bars on all backbone rows, the conclusion that appearance features provide 'limited or negative gains' is not fully supported. Please state the split explicitly and provide error bars for every configuration in Table I.
  3. [IV-F, Table II] The CAD evaluation protocol is under-specified. The text says pairwise predictions are mapped to group structures and group activities via majority voting, but the mapping is not described, and 'Membership Acc' and 'Social Activity Acc' are not defined in Section IV-D. Furthermore, the claim that the method is 'competitive' with fully supervised methods is strained: ours achieves 79.8% membership and 52.6% social activity accuracy, versus 83.0% and 69.0% for I3D-SA-GAT [learn2cluster]. The gap on social activity is substantial. Please clarify the mapping and metrics, and qualify the claim appropriately.
  4. [IV-G] The zero-shot lawnmower evaluation is described as demonstrating 'robust zero shot transfer,' but the reported Stage 2 macro F1 is 0.51. The paper reports high recall for walking and standing classes, but most sitting interactions are misclassified as walking, which is a substantial error mode for social navigation. Please report per-class precision/recall or a confusion matrix, state the class distribution in the lawnmower data, and temper the 'robust' characterization accordingly.
minor comments (5)
  1. [References] References [29] and [30] are the same paper (Yuan et al., Spatio-temporal dynamic inference network for group activity recognition). Please remove the duplicate.
  2. [Equations (3)-(7)] The symbol \bar{h} is used for mean bounding-box height in Eqs. (3) and (5) and for mean horizontal optical flow in Eq. (7). Please use distinct symbols to avoid confusion.
  3. [Fig. 1] The label '1 2 9D Sym' below the flow feature block is unclear. Please spell out the symmetrization operation in the caption or text.
  4. [IV-B] The frame sampling interval is said to be 'an interval' but the actual value is not given. Please specify the interval.
  5. [Table II] The abbreviation 'GT' is not defined, and it is unclear what 'ARG' in the group/individuals rows refers to. Please clarify the baselines.

Circularity Check

0 steps flagged

No significant circularity: the paper's results are measured empirical outcomes, not derived from fitted inputs or self-citation chains.

full rationale

The paper's central claim—that simple geometric and motion cues provide an efficient basis for pairwise interaction perception—is supported by direct empirical evaluation, not by a derivation that reduces to its own inputs. Stage 1 features (Eqs. 1-2) and Stage 2 geometric/motion features (Eqs. 3-9) are explicit hand-defined inputs, and the reported 84.3% accuracy on JRDB (Table I) is a measured ablation result, not a parameter fitted from the labels and then renamed as a prediction. The comparison against appearance-based backbones is likewise an empirical ablation, and the CAD and zero-shot lawnmower experiments provide external, in-distribution and out-of-distribution tests. No self-citations appear in the reference list, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via citation: the geometric/motion design is described directly in the method section. The limitations section acknowledges dependence on reliable person detection and tracking, and Section IV-G reports only Stage 1 precision (96.5%) without recall; this is an evaluation gap and a deployment-risk concern, not a circular step, because omitting a metric does not make the reported accuracy equivalent to the model's inputs. The label-space reduction is justified by dataset label statistics, which is in-distribution design rather than circular reasoning. Overall, the derivation chain is self-contained and evidence-based, so the circularity score is 0.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The paper is an empirical ML system; its free parameters are the trained weights and unspecified hyperparameters. The axioms are domain assumptions about the sufficiency of coarse pairwise interaction and low-level cues; none are formal mathematical axioms. No invented entities.

free parameters (6)
  • Stage 1 feature reweighting vectors b, s, w = learned (init b=0, s=w=1)
    Eq. (1) reweights the 7D geometric features; learned on JRDB, so the detection output depends on fitted values.
  • Stage 1 MLP weights (32→16→2) = learned
    Binary interaction detector network; trained on JRDB.
  • Stage 2 relation network weights = learned
    1152→512→256→128→3 classifier; trained on JRDB/CAD.
  • Detection threshold θ = not reported
    Stage 1 uses a 'lower threshold θ' to favor recall; the value is absent, so the detection operating point is underspecified.
  • Focal loss hyperparameters α, γ = not specified
    Focal loss is used for class imbalance but α/γ are not reported.
  • Frame sampling interval = not specified
    Frames are sampled at an interval to reduce temporal redundancy; the interval value is not given.
axioms (6)
  • domain assumption Pairwise interactions are a minimal yet sufficient perceptual unit for mobile robot social understanding.
    Central framing from the Introduction; if false, the entire task definition and evaluation are misdirected.
  • domain assumption Bounding-box geometry and optical flow contain enough signal to discriminate walking/standing/sitting together.
    The method's premise; tested only in-distribution on JRDB and with weak zero-shot sitting performance on the lawnmower.
  • domain assumption JRDB intra-group annotations can be reduced to three coarse classes without losing robot-relevant structure.
    Section IV-A justifies the reduction by label dependencies, but it remains a modeling choice that affects all reported numbers.
  • domain assumption Frozen EfficientNet features are an adequate appearance representation; training the backbone is unnecessary.
    Section IV-C freezes the backbone; if fine-tuning were needed, the appearance ablation would not be a fair comparison.
  • domain assumption Motion compensation on the lawnmower removes ego-motion sufficiently for Farneback flow to reflect human motion.
    Section IV-G: the sitting→walking confusion shows this assumption is violated in the zero-shot deployment setting.
  • domain assumption Reliable person detection and tracking is available upstream.
    Stated in the Limitations section; if detection/tracking fails, the pair proposal stage cannot work.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A Pairwise Human-Human Interaction Detection and Recognition Framework for Mobile Service Robots." pith.science (2026). https://pith.science/paper/F2GG4S5C

@misc{pith2026260222346,
  author       = {Pith},
  title        = {Pith review of: A Pairwise Human-Human Interaction Detection and Recognition Framework for Mobile Service Robots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F2GG4S5C}},
  note         = {Machine review of arXiv:2602.22346}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Autonomous mobile service robots, such as lawnmowers or cleaning robots, operating in human-populated environments need to reason about human-human interactions to support safe and socially aware navigation. For such systems, interaction understanding is not primarily a fine-grained recognition problem, but a perception problem under limited sensing quality and computational resources. Many existing approaches focus on holistic group activity recognition, often relying on complex and computationally expensive models that are not well suited for mobile robotic platforms. In this work, we argue that pairwise human interactions constitute a minimal yet sufficient perceptual unit for robot-centric social understanding. We study the problem of identifying interacting person pairs and classifying coarse-grained interaction behaviors sufficient for downstream group-level reasoning and robot decision-making. To this end, we adopt a two-stage framework in which candidate interacting pairs are first identified using lightweight geometric and motion cues, and interaction types are subsequently classified using a relation network. We evaluate the proposed approach on the JRDB dataset, where it achieves competitive performance with reduced computational cost and model size compared to appearance-based methods. Additional experiments on the Collective Activity Dataset (CAD) and zero-shot evaluation on a lawnmower-collected dataset further demonstrate the generalizability of the proposed framework. These results suggest that simple geometric and motion cues provide a practical and efficient basis for interaction-aware perception in mobile service robots. Code is released.

Figures

Figures reproduced from arXiv: 2602.22346 by Iolanda Leite, Mengyu Liang, Sarah Gillet.

Figure 1
Figure 1. Figure 1: Overview of the proposed two-stage pairwise interaction recognition framework: Stage 1 performs interaction detection using a 7D geometric feature vector derived from bounding box configurations, producing candidate interacting person pairs. Stage 2 classifies coarse-grained interaction types by combining frozen visual appearance features extracted by EfficientNet with geometric–motion features computed fr… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative zero shot interaction recognition results on data collected from a mobile lawnmower platform (a) The lawnmower moves rapidly through a group of pedestrians, inducing strong ego-motion and viewpoint changes, which lead to misclassifications of interaction types. (b) The lawnmower remains stationary, resulting in stable observations and correct interaction recognition most person pairs. (c) The l… view at source ↗

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

31 extracted references · 5 canonical work pages

  1. [1]

    Farzaneh Askari, Cyril Yared, Rohit Ramaprasad, Devin Garg, Anjun Hu, and James J. Clark. Video interac- tion recognition using an attention augmented relational network and skeleton data. In2024 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3225–3234, June 2024. doi: 10.1109/CVPRW63382.2024.00328

    arXiv 2024

  2. [2]

    Social scene understand- ing: End-to-end multi-person action localization and col- lective activity recognition

    Timur Bagautdinov, Alexandre Alahi, Franc ¸ois Fleuret, Pascal Fua, and Silvio Savarese. Social scene understand- ing: End-to-end multi-person action localization and col- lective activity recognition. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3425–3434, July 2017. doi: 10.1109/CVPR.2017.365

    work page doi:10.1109/cvpr.2017.365 2017

  3. [3]

    Recent trends in crowd analysis: A review.Machine Learning with Applications, 4:100023, 2021

    Mounir Bendali-Braham, Jonathan Weber, Germain Forestier, Lhassane Idoumghar, and Pierre-Alain Muller. Recent trends in crowd analysis: A review.Machine Learning with Applications, 4:100023, 2021. ISSN 2666-8270. doi: https://doi.org/10.1016/j.mlwa.2021. 100023. URL https://www.sciencedirect.com/science/ article/pii/S2666827021000049

    work page doi:10.1016/j.mlwa.2021 2021

  4. [4]

    What are they doing? : Collective activity classification using spatio-temporal relationship among people

    Wongun Choi, Khuram Shahid, and Silvio Savarese. What are they doing? : Collective activity classification using spatio-temporal relationship among people. In 2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV Workshops, pages 1282–1289, Sep. 2009. doi: 10.1109/ICCVW.2009.5457461

    arXiv 2009

  5. [5]

    Learning context for collective activity recognition

    Wongun Choi, Khuram Shahid, and Silvio Savarese. Learning context for collective activity recognition. In CVPR 2011, pages 3273–3280, June 2011. doi: 10.1109/ CVPR.2011.5995707

    arXiv 2011

  6. [6]

    Joint learning of social groups, individuals action and sub-group activities in videos, 2020

    Mahsa Ehsanpour, Alireza Abedin, Fatemeh Saleh, Javen Shi, Ian Reid, and Hamid Rezatofighi. Joint learning of social groups, individuals action and sub-group activities in videos, 2020. URL https://arxiv.org/abs/2007.02632

    Pith/arXiv arXiv 2020

  7. [7]

    Two-frame motion estimation based on polynomial expansion

    Gunnar Farneb ¨ack. Two-frame motion estimation based on polynomial expansion. volume 2749, pages 363– 370, 06 2003. ISBN 978-3-540-40601-3. doi: 10.1007/ 3-540-45103-X 50

    2003

  8. [8]

    Deep residual learning for image recognition, 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. URL https://arxiv.org/abs/1512.03385

    Pith/arXiv arXiv 2015

  9. [9]

    Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam

    Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision appli- cations, 2017. URL https://arxiv.org/abs/1704.04861

    Pith/arXiv arXiv 2017

  10. [10]

    Ibrahim, Srikanth Muralidharan, Zhiwei Deng, Arash Vahdat, and Greg Mori

    Mostafa S. Ibrahim, Srikanth Muralidharan, Zhiwei Deng, Arash Vahdat, and Greg Mori. A hierarchical deep temporal model for group activity recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1971–1980, June 2016. doi: 10.1109/CVPR.2016.217

    work page doi:10.1109/cvpr.2016.217 2016

  11. [11]

    Real-time trajectory-based social group detection, 2023

    Simindokht Jahangard, Munawar Hayat, and Hamid Rezatofighi. Real-time trajectory-based social group detection, 2023. URL https://arxiv.org/abs/2304.05678

    Pith/arXiv arXiv 2023

  12. [12]

    Jrdb-social: A multifaceted robotic dataset for understanding of context and dynamics of human interactions within social groups

    Simindokht Jahangard, Zhixi Cai, Shiki Wen, and Hamid Rezatofighi. Jrdb-social: A multifaceted robotic dataset for understanding of context and dynamics of human interactions within social groups. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 22087–22097, June 2024. doi: 10.1109/CVPR52733.2024.02085

    arXiv 2024

  13. [13]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C.J. Burges, L. Bottou, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 25. Curran Associates, Inc., 2012. URL https://proceedings.neurips.cc/paper files/paper/2012/ file/c399862d3b...

    2012

  14. [14]

    Robi- novitch, and Greg Mori

    Tian Lan, Yang Wang, Weilong Yang, Stephen N. Robi- novitch, and Greg Mori. Discriminative latent models for recognizing contextual group activities.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 34(8):1549–1562, Aug 2012. ISSN 1939-3539. doi: 10.1109/TPAMI.2011.228

    work page doi:10.1109/tpami.2011.228 2012

  15. [16]

    Two-person graph convolutional network for skeleton-based human interaction recogni- tion.IEEE Trans

    Zhengcen Li, Yueran Li, Linlin Tang, Tong Zhang, and Jingyong Su. Two-person graph convolutional network for skeleton-based human interaction recogni- tion.IEEE Trans. Cir. and Sys. for Video Technol., 33(7):3333–3342, July 2023. ISSN 1051-8215. doi: 10.1109/TCSVT.2022.3232373. URL https://doi.org/10. 1109/TCSVT.2022.3232373

    arXiv 2023

  16. [17]

    Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C. Kot. Ntu rgb+d 120: A large- scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684–2701, Oct 2020. ISSN 1939-

    2020

  17. [18]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. URL https://arxiv.org/abs/ 1711.05101

    Pith/arXiv arXiv 2019

  18. [19]

    Structured learning of human inter- actions in tv shows.IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, 34(12):2441–2453, Dec

    Alonso Patron-Perez, Marcin Marszalek, Ian Reid, and Andrew Zisserman. Structured learning of human inter- actions in tv shows.IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, 34(12):2441–2453, Dec

  19. [20]

    Mauricio Perez, Jun Liu, and Alex C. Kot. Interaction relational network for mutual action recognition.IEEE Transactions on Multimedia, 24:366–376, 2022. ISSN 1941-0077. doi: 10.1109/TMM.2021.3050642

    arXiv 2022

  20. [21]

    stagnet: An attentive semantic rnn for group activity recognition

    Mengshi Qi, Jie Qin, Annan Li, Yunhong Wang, Jiebo Luo, and Luc Van Gool. stagnet: An attentive semantic rnn for group activity recognition. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors,Computer Vision – ECCV 2018, pages 104–120, Cham, 2018. Springer International Publishing. ISBN 978-3-030-01249-6

    2018

  21. [22]

    M. S. Ryoo and J. K. Aggarwal. UT- Interaction Dataset, ICPR contest on Semantic Description of Human Activities (SDHA). http://cvrc.ece.utexas.edu/SDHA2010/Human Interaction.html, 2010

    2010

  22. [23]

    Very deep convolutional networks for large-scale image recognition,

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition,

  23. [24]

    Mingxing Tan and Quoc V . Le. Efficientnet: Rethinking model scaling for convolutional neural networks, 2020. URL https://arxiv.org/abs/1905.11946

    Pith/arXiv arXiv 2020

  24. [25]

    Angelique Taylor and Laurel D. Riek. Regroup: A robot- centric group detection and tracking system. In2022 17th ACM/IEEE International Conference on Human-Robot Interaction (HRI), pages 412–421, March 2022. doi: 10. 1109/HRI53351.2022.9889634

    arXiv 2022

  25. [26]

    Non-local neural networks

    Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7794–7803, 2018. doi: 10.1109/ CVPR.2018.00813

    arXiv 2018

  26. [27]

    A review of human-object interaction detection, 2025

    Yuxiao Wang, Yu Lei, Li Cui, Weiying Xue, Qi Liu, and Zhenao Wei. A review of human-object interaction detection, 2025. URL https://arxiv.org/abs/2408.10641

    Pith/arXiv arXiv 2025

  27. [28]

    Learning actor relation graphs for group activity recognition

    Jianchao Wu, Limin Wang, Li Wang, Jie Guo, and Gangshan Wu. Learning actor relation graphs for group activity recognition. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9956–9966, June 2019. doi: 10.1109/CVPR.2019.01020

    arXiv 2019

  28. [30]

    Spatio- temporal dynamic inference network for group activity recognition

    Hangjie Yuan, Dong Ni, and Mang Wang. Spatio- temporal dynamic inference network for group activity recognition. In2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7456–7465, Oct 2021. doi: 10.1109/ICCV48922.2021.00738

    arXiv 2021

  29. [2012]

    doi: 10.1109/TPAMI.2012.24

    ISSN 1939-3539. doi: 10.1109/TPAMI.2012.24

    work page doi:10.1109/tpami.2012.24 1939

  30. [2015]

    URL https://arxiv.org/abs/1409.1556

    Pith/arXiv arXiv

  31. [3539]

    doi: 10.1109/TPAMI.2019.2916873

    arXiv 2019

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.