arxiv: 2604.02616 · v1 · submitted 2026-04-03 · 💻 cs.CV

Recognition: 2 theorem links

· Lean Theorem

Unlocking Multi-Site Clinical Data: A Federated Approach to Privacy-First Child Autism Behavior Analysis

Guangyu Sun , Wenhan Wu , Zhishuai Guo , Ziteng Wang , Pegah Khosravi , Chen Chen

Authors on Pith no claims yet

Pith reviewed 2026-05-13 20:00 UTC · model grok-4.3

classification 💻 cs.CV

keywords federated learningautism behavior recognitionskeletal pose abstractionprivacy preservationmulti-site clinical datachild autism analysispose-based recognition

0 comments

The pith

Federated learning with skeletal poses enables accurate autism behavior recognition across clinical sites without sharing raw patient data.

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

The paper presents the first application of federated learning to pose-based recognition of autistic behaviors in children. It addresses barriers from privacy rules like HIPAA and limited data at individual clinics by keeping all raw video and pose information local while still training a shared model. A two-layer approach first converts RGB videos to skeletal poses to strip identifying visuals, then applies federated learning so each site trains on its own data and shares only model updates. Experiments on the MMASD benchmark show this yields higher accuracy than standard federated baselines and supports both generalized patterns and site-specific tailoring for early clinical assessment.

Core claim

Our framework employs a two-layer privacy protection mechanism: utilizing human skeletal abstraction to remove identifiable visual information from the raw RGB videos and FL to ensure sensitive pose data remains within the clinic. This approach leverages distributed clinical data to learn generalized representations while providing the flexibility for site-specific personalization. Experimental results on the MMASD benchmark demonstrate that our framework achieves high recognition accuracy, outperforming traditional federated baselines and providing a robust, privacy-first solution for multi-site clinical analysis.

What carries the argument

Two-layer privacy mechanism of human skeletal pose abstraction from RGB videos followed by federated learning that keeps all pose data local at each clinic.

If this is right

Multiple clinical sites can jointly improve behavior recognition models without ever pooling or moving patient videos.
Each site can further adapt the shared model to its own patient distribution while still using knowledge from other sites.
Objective, data-driven autism assessment becomes feasible at scale even when data scarcity and privacy rules block centralized training.

Where Pith is reading between the lines

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

The same skeletal-plus-federated pattern could extend to other video-based pediatric or neurological assessments where raw footage must stay private.
Adding differential privacy on the model updates might close remaining leakage risks from shared gradients.
Performance on larger numbers of heterogeneous sites would test whether the abstraction step remains sufficient as data variety grows.

Load-bearing premise

That converting raw RGB videos to human skeletal poses removes all identifiable visual information while still preserving the behavior patterns needed for accurate autism recognition.

What would settle it

A test showing that individual children can be reliably re-identified from the extracted skeletal pose sequences alone would invalidate the privacy protection of the first layer.

Figures

Figures reproduced from arXiv: 2604.02616 by Chen Chen, Guangyu Sun, Pegah Khosravi, Wenhan Wu, Zhishuai Guo, Ziteng Wang.

**Figure 1.** Figure 1: Overview of the proposed Two-Layer Privacy framework. The first layer achieves privacy via skeletal abstraction, filtering out raw biometric identifiers from videos. This is done at each clinical site for their local video data (a demonstration is shown for ‘Clinical Site A’ in this figure). The second layer employs decentralized optimization (FL) using the efficient FreqMixFormer backbone to maintain data… view at source ↗

**Figure 2.** Figure 2: Visualization of 3D skeletal data from the MMASD benchmark. (a) shows representative sequences for Roboticassisted therapy, Rhythm-based activities, and Yoga-based poses. (b) illustrates the 3D joint structure. This abstraction preserves kinetic motion for behavior analysis while raw biometric identifiers are removed. 4.1. Experimental Setup Dataset: MMASD Benchmark. We evaluate our framework using the … view at source ↗

**Figure 3.** Figure 3: Performance evolution across 30 communication rounds for different clinical themes. The curves demonstrate the convergence characteristics of standard FL, parameter-wise PFL, and our adaptive personalization approach. APFL (red) consistently achieves superior stability and final recognition accuracy across all therapeutic domains. 0 5 10 15 20 25 30 Communication Round 0.0 0.2 0.4 0.6 0.8 1.0 Mixin g P a … view at source ↗

**Figure 4.** Figure 4: Evolution of the adaptive mixing parameter α across different clinical themes. The parameter is initialized with a low value and adaptively increases, indicating that the model initially prioritizes global knowledge synthesized from the collaborative network before gradually incorporating site-specific behavioral nuances from local data. be degraded by clinical heterogeneity, our adaptive personalized app… view at source ↗

read the original abstract

Automated recognition of autistic behaviors in children is essential for early intervention and objective clinical assessment. However, the development of robust models is severely hindered by strict privacy regulations (e.g., HIPAA) and the sensitive nature of pediatric data, which prevents the centralized aggregation of clinical datasets. Furthermore, individual clinical sites often suffer from data scarcity, making it difficult to learn generalized behavior patterns or tailor models to site-specific patient distributions. To address these challenges, we observe that Federated Learning (FL) can decouple model training from raw data access, enabling multi-site collaboration while maintaining strict data residency. In this paper, we present the first study exploring Federated Learning for pose-based child autism behavior recognition. Our framework employs a two-layer privacy protection mechanism: utilizing human skeletal abstraction to remove identifiable visual information from the raw RGB videos and FL to ensure sensitive pose data remains within the clinic. This approach leverages distributed clinical data to learn generalized representations while providing the flexibility for site-specific personalization. Experimental results on the MMASD benchmark demonstrate that our framework achieves high recognition accuracy, outperforming traditional federated baselines and providing a robust, privacy-first solution for multi-site clinical analysis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper applies standard federated averaging after skeletal pose extraction to enable multi-site autism behavior recognition without sharing raw video, but the abstract gives no numbers or ablations to support the accuracy claims.

read the letter

The main takeaway is that this is the first reported use of federated learning on pose data for child autism behavior recognition. The setup uses human skeletal abstraction to strip visual identifiers from RGB videos, then runs federated averaging across clinical sites so each keeps its own data while still training a shared model with some site-specific adaptation. That combination directly targets the real constraints of HIPAA and small per-site datasets for a rare pediatric condition. The MMASD benchmark is a sensible choice for testing, and the framing of privacy plus generalization is straightforward and practical. Standard FedAvg is applied without new math, which fits an application paper and keeps the method reproducible from existing tools. The stress-test note on pose leakage is worth checking in the full text, but the abstract itself does not contradict the basic logic of the two-layer approach. The soft spot is the missing evidence. The abstract states high accuracy and outperformance over federated baselines yet supplies no numeric results, no error bars, no baseline implementation details, and no ablation comparing pose input to raw video. There are also no re-identification tests or checks on whether subtle cues like hand movements survive the pose estimator. Without those, the performance and privacy claims cannot be evaluated. The paper is aimed at applied researchers in medical computer vision or privacy-preserving ML who need templates for multi-site pediatric work. A reader could adapt the pipeline, but would need the full experimental section to decide if the results hold. I would send it to peer review because the application is timely and the privacy problem is genuine; referees can require the missing metrics and ablations to make the contribution usable.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes a two-layer privacy-preserving framework for multi-site child autism behavior recognition: skeletal pose abstraction from RGB videos to remove identifiable visual information, combined with federated learning (standard averaging) to keep pose data local at clinical sites. It claims to be the first such study and asserts that experiments on the MMASD benchmark show high recognition accuracy that outperforms traditional federated baselines while enabling site-specific personalization.

Significance. If the experimental claims are substantiated with quantitative results, ablations, and privacy metrics, the work would address a genuine barrier in pediatric clinical ML by enabling collaborative training across HIPAA-regulated sites without centralizing raw video. The approach is conceptually straightforward and leverages existing tools (pose estimation + FL), but its impact hinges on demonstrating that pose abstraction neither leaks identity nor discards autism-relevant cues.

major comments (3)

[Abstract] Abstract: the claim that the framework 'achieves high recognition accuracy, outperforming traditional federated baselines' supplies no numeric accuracies, standard deviations, ablation tables, or baseline implementation details (e.g., which FL variants or pose estimators were used), rendering the central empirical result unverifiable.
[Abstract] Abstract / Methods: the core assumption that skeletal pose abstraction 'removes all identifiable visual information' while 'preserving the behavior patterns needed for accurate autism recognition' is unsupported by any re-identification attack results, RGB-vs-pose accuracy ablation, or analysis of retained cues (e.g., hand stereotypies or gait signatures); this directly affects both the privacy and utility claims.
[Experimental results] Experimental results: no site-specific data distribution statistics, cross-site generalization metrics, or details on how the MMASD benchmark was partitioned across sites are provided, which are required to substantiate the multi-site collaboration benefit.

minor comments (1)

[Abstract] Abstract: consider adding one sentence on the number of clinical sites and total video hours in MMASD to ground the multi-site claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point-by-point below, providing clarifications from the manuscript and indicating revisions where the presentation can be strengthened without misrepresenting the work.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that the framework 'achieves high recognition accuracy, outperforming traditional federated baselines' supplies no numeric accuracies, standard deviations, ablation tables, or baseline implementation details (e.g., which FL variants or pose estimators were used), rendering the central empirical result unverifiable.

Authors: We agree that the abstract would be more informative with explicit numbers. The full manuscript reports these results in Section 4 (Experimental Results), including per-class accuracies with standard deviations across multiple runs, ablation tables comparing FL variants, and implementation details (OpenPose for skeletal extraction and standard FedAvg for aggregation). To improve verifiability at a glance, we will revise the abstract to include the key numeric outcomes (e.g., overall accuracy and margin over baselines). revision: yes
Referee: [Abstract] Abstract / Methods: the core assumption that skeletal pose abstraction 'removes all identifiable visual information' while 'preserving the behavior patterns needed for accurate autism recognition' is unsupported by any re-identification attack results, RGB-vs-pose accuracy ablation, or analysis of retained cues (e.g., hand stereotypies or gait signatures); this directly affects both the privacy and utility claims.

Authors: The manuscript grounds the privacy claim in the fact that skeletal abstraction discards RGB pixels entirely (no faces, clothing, or backgrounds remain), which is a standard argument in pose-based privacy literature. However, we did not perform explicit re-identification attacks or direct RGB-vs-pose ablations because raw RGB video cannot be centralized or even accessed for comparison under the multi-site HIPAA constraints that motivate the work. In revision we will add a dedicated privacy subsection with a simulated re-identification experiment on the available pose sequences and a qualitative discussion of retained kinematic cues (e.g., repetitive hand movements) supported by clinical references. A full RGB ablation remains infeasible without violating data-residency rules. revision: partial
Referee: [Experimental results] Experimental results: no site-specific data distribution statistics, cross-site generalization metrics, or details on how the MMASD benchmark was partitioned across sites are provided, which are required to substantiate the multi-site collaboration benefit.

Authors: We will expand the Experimental Setup subsection to include per-site sample counts, class distributions, and any available demographic statistics from the MMASD benchmark. The benchmark was partitioned by originating clinical site to simulate realistic data silos; we will explicitly state this partitioning strategy and add leave-one-site-out cross-site generalization results to quantify the benefit of federated training over local-only models. revision: yes

Circularity Check

0 steps flagged

No circularity: standard FL and pose estimation with independent experimental results

full rationale

The paper describes a two-layer framework combining skeletal pose abstraction from RGB videos with federated learning for multi-site autism behavior recognition, reporting empirical accuracy gains on the MMASD benchmark over traditional federated baselines. No equations, parameter fittings, or derivations are presented that reduce claimed performance to inputs defined inside the paper. The approach invokes established FL averaging and pose estimation techniques without self-citation chains, uniqueness theorems, or ansatzes that would make the central result tautological. The derivation remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Only the abstract is available, so the ledger records the core assumptions stated there. The framework depends on the unproven premise that skeletal abstraction preserves diagnostic behavior signals while eliminating identity.

axioms (2)

domain assumption Skeletal pose abstraction removes identifiable visual information from RGB videos while retaining autism-relevant movement patterns
Invoked in the two-layer privacy mechanism description.
domain assumption Federated averaging across clinical sites produces a model that generalizes better than local-only training
Central premise of the federated learning component.

pith-pipeline@v0.9.0 · 5519 in / 1264 out tokens · 56453 ms · 2026-05-13T20:00:55.763877+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Our framework employs a two-layer privacy protection mechanism: utilizing human skeletal abstraction to remove identifiable visual information from the raw RGB videos and FL to ensure sensitive pose data remains within the clinic.
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We utilize the efficient FreqMixFormer architecture, incorporating frequency-aware attention to ensure robust behavioral analysis across distributed clinical nodes.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages · 1 internal anchor

[1]

Discrete cosine transform.IEEE transactions on Computers, 100(1): 90–93, 1974

Nasir Ahmed, T Natarajan, and Kamisetty R Rao. Discrete cosine transform.IEEE transactions on Computers, 100(1): 90–93, 1974. 4

work page 1974
[2]

Federated Learning with Personalization Layers

Manoj Ghuhan Arivazhagan, Vinay Aggarwal, Aaditya Ku- mar Singh, and Sunav Choudhary. Federated learning with personalization layers.arXiv preprint arXiv:1912.00818,

work page internal anchor Pith review arXiv 1912
[3]

Automated analysis of stereotypical movements in videos of children with autism spectrum disorder.JAMA network open, 7(9): e2432851, 2024

Tal Barami, Liora Manelis-Baram, Hadas Kaiser, Michal Ilan, Aviv Slobodkin, Ofri Hadashi, Dor Hadad, Danel Wais- sengreen, Tanya Nitzan, Idan Menashe, et al. Automated analysis of stereotypical movements in videos of children with autism spectrum disorder.JAMA network open, 7(9): e2432851, 2024. 1, 2

work page 2024
[4]

Skeleton-based action recogni- tion with shift graph convolutional network

Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recogni- tion with shift graph convolutional network. InCVPR, pages 183–192, 2020. 2

work page 2020
[5]

Language-assisted deep learning for autistic behaviors recognition.Smart Health, 32:100444, 2024

Andong Deng, Taojiannan Yang, Chen Chen, Qian Chen, Leslie Neely, and Sakiko Oyama. Language-assisted deep learning for autistic behaviors recognition.Smart Health, 32:100444, 2024. 1

work page 2024
[6]

Adaptive personalized fed- erated learning.arXiv preprint arXiv:2003.13461,

Yuyang Deng, Mohammad Mahdi Kamani, and Mehrdad Mahdavi. Adaptive personalized federated learning.arXiv preprint arXiv:2003.13461, 2020. 2, 3, 5, 6

work page arXiv 2003
[7]

Per- sonalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach

Alireza Fallah, Aryan Mokhtari, and Asuman Ozdaglar. Per- sonalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. InNeurIPS, 2020. 3

work page 2020
[8]

Federated action recognition for smart worker assistance using fastpose, 2025

Vinit Hegiste, Vidit Goyal, Tatjana Legler, and Martin Ruskowski. Federated action recognition for smart worker assistance using fastpose, 2025. 3

work page 2025
[9]

Mmasd: A multimodal dataset for autism intervention analysis

Jicheng Li, Vuthea Chheang, Pinar Kullu, Eli Brignac, Zhang Guo, Anjana Bhat, Kenneth E Barner, and Roghayeh Leila Barmaki. Mmasd: A multimodal dataset for autism intervention analysis. InProceedings of the 25th International Conference on Multimodal Interaction, pages 397–405, 2023. 1, 2, 4, 5

work page 2023
[10]

Federated optimiza- tion in heterogeneous networks

Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimiza- tion in heterogeneous networks. InProceedings of Machine Learning and Systems (MLSys), pages 429–450, 2020. 3, 4, 6

work page 2020
[11]

Ditto: Fair and robust federated learning through per- sonalization

Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through per- sonalization. InICML, pages 6357–6368, 2021. 3

work page 2021
[12]

Fedbn: Federated learning on non-iid features via local batch normalization

Xiaoxiao Li, Meirui Jiang, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization. InICLR, 2021. 2, 3, 5, 6

work page 2021
[13]

Disentangling and unifying graph con- volutions for skeleton-based action recognition

Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph con- volutions for skeleton-based action recognition. InCVPR, pages 143–152, 2020. 2

work page 2020
[14]

Personalized feder- ated learning with adaptive batchnorm for healthcare.IEEE Transactions on Big Data, 10(6):915–925, 2024

Wang Lu, Jindong Wang, Yiqiang Chen, Xin Qin, Renjun Xu, Dimitrios Dimitriadis, and Tao Qin. Personalized feder- ated learning with adaptive batchnorm for healthcare.IEEE Transactions on Big Data, 10(6):915–925, 2024. 3

work page 2024
[15]

Communication- Efficient Learning of Deep Networks from Decentralized Data

Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication- Efficient Learning of Deep Networks from Decentralized Data. InProceedings of the 20th International Conference on Artificial Intelligence and Statistics, pages 1273–1282. PMLR, 2017. 2, 3, 4, 6

work page 2017
[16]

Efraimidis, and Despina Elisabeth Filippidou

Nikolaos Pavlidis, Vasileios Perifanis, Eleni Briola, Christos-Chrysanthos Nikolaidis, Eleftheria Katsiri, Pav- los S Efraimidis, and Despina Elisabeth Filippidou. Fed- erated anomaly detection for early-stage diagnosis of autism spectrum disorders using serious game data.arXiv preprint arXiv:2410.20003, 2024. 3

work page arXiv 2024
[17]

Early detection of autism using dig- ital behavioral phenotyping.Nature Medicine, 29(10):2489– 2497, 2023

Sam Perochon, J Matias Di Martino, Kimberly LH Carpen- ter, Scott Compton, Naomi Davis, Brian Eichner, Steven Espinosa, Lauren Franz, Pradeep Raj Krishnappa Babu, Guillermo Sapiro, et al. Early detection of autism using dig- ital behavioral phenotyping.Nature Medicine, 29(10):2489– 2497, 2023. 1, 2

work page 2023
[18]

Mmasd+: A novel dataset for privacy-preserving behavior analysis of children with autism spectrum disorder.arXiv preprint arXiv:2408.15077, 2024

Pavan Uttej Ravva, Behdokht Kiafar, Pinar Kullu, Jicheng Li, Anjana Bhat, and Roghayeh Leila Barmaki. Mmasd+: A novel dataset for privacy-preserving behavior analysis of children with autism spectrum disorder.arXiv preprint arXiv:2408.15077, 2024. 2

work page arXiv 2024
[19]

Imitasd: Imitation assessment model for children with autism based on human pose estimation.Mathematics, 12(21):3438, 2024

Hany Said, Khaled Mahar, Shaymaa E Sorour, Ahmed Elsheshai, Ramy Shaaban, Mohamed Hesham, Mustafa Khadr, Youssef A Mehanna, Ammar Basha, and Fahima A Maghraby. Imitasd: Imitation assessment model for children with autism based on human pose estimation.Mathematics, 12(21):3438, 2024. 1, 2

work page 2024
[20]

Athmar NM Shamhan, Marwa Qaraqe, and Dena Al-Thani. Advancements in automated assessment and diagnosis of autism spectrum disorder through multi-modality sensing technologies: Survey of the last decade.IEEE Transactions on Cognitive and Developmental Systems, 2025. 1, 2

work page 2025
[21]

On the feasibility of federated learning for neurodevelopmen- tal disorders: Asd detection use-case

Hala Shamseddine, Safa Otoum, and Azzam Mourad. On the feasibility of federated learning for neurodevelopmen- tal disorders: Asd detection use-case. InGLOBECOM 2022-2022 IEEE Global Communications Conference, pages 1121–1127. IEEE, 2022. 3

work page 2022
[22]

Federated learning in medicine: facilitating multi- institutional collaborations without sharing patient data.Sci- entific Reports, 10(1):12598, 2020

Micah J Sheller, Brandon Edwards, G Anthony Reina, Ja- son Martin, Sarthak Pati, Aikaterini Kotrotsou, Mikhail Milchenko, Weilin Xu, Daniel Marcus, Rivka R Colen, et al. Federated learning in medicine: facilitating multi- institutional collaborations without sharing patient data.Sci- entific Reports, 10(1):12598, 2020. 1, 3

work page 2020
[23]

World-grounded human motion recovery via gravity-view coordinates

Zehong Shen, Huaijin Pi, Yan Xia, Zhi Cen, Sida Peng, Zechen Hu, Hujun Bao, Ruizhen Hu, and Xiaowei Zhou. World-grounded human motion recovery via gravity-view coordinates. InSIGGRAPH Asia, 2024. 4

work page 2024
[24]

Skeleton-based action recognition with multi-stream adap- tive graph convolutional networks.IEEE Transactions on Image Processing, 29:9532–9545, 2020

Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adap- tive graph convolutional networks.IEEE Transactions on Image Processing, 29:9532–9545, 2020. 2

work page 2020
[25]

To- wards personalized federated learning.IEEE transactions on neural networks and learning systems, 34(12):9587–9603,

Alysa Ziying Tan, Han Yu, Lizhen Cui, and Qiang Yang. To- wards personalized federated learning.IEEE transactions on neural networks and learning systems, 34(12):9587–9603,

work page
[26]

Frequency guidance matters: Skeletal ac- tion recognition by frequency-aware mixed transformer

Wenhan Wu, Ce Zheng, Zihao Yang, Chen Chen, Srijan Das, and Aidong Lu. Frequency guidance matters: Skeletal ac- tion recognition by frequency-aware mixed transformer. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1632–1641, 2024. 2, 3, 4, 6

work page 2024
[27]

Federated learning for healthcare informatics.Journal of healthcare informatics research, 5 (1):1–19, 2021

Jie Xu, Benjamin S Glicksberg, Chang Su, Peter Walker, Jiang Bian, and Fei Wang. Federated learning for healthcare informatics.Journal of healthcare informatics research, 5 (1):1–19, 2021. 3

work page 2021
[28]

Spatial tempo- ral graph convolutional networks for skeleton-based action recognition

Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial tempo- ral graph convolutional networks for skeleton-based action recognition. InAAAI, 2018. 2 10

work page 2018