Pith. sign in

REVIEW 4 major objections 6 minor 33 references

Interact with me: Joint Egocentric Forecasting of Intent to Interact, Attitude and Social Actions

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper proposes that a compact graph network, SocialEgoNet, can jointly forecast a person's intent to interact, their attitude, and their next action from one second of egocentric skeleton video.

desk verdict A genuinely new joint forecasting task and a useful annotated egocentric dataset, but the attitude label is a deterministic relabeling of the action taxonomy, so the headline accuracy partly measures action discrimination rather than an independent social signal. read the letter →

arxiv 2412.16698 v3 pith:ADSC5OGL submitted 2024-12-21 cs.CV cs.HC

Tongfei Bian , Yiming Ma , Mathieu Chollet , Victor Sanchez , Tanaya Guha This is my paper · ORCID
classification cs.CVcs.HC
keywords egocentricvisionhuman-robotinteractionintentpredictionskeleton-basedactionrecognitionmultitasklearninggraphconvolutionalnetworksocialforecasting
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

This paper proposes a new task for social robotics: jointly forecasting, from one second of egocentric video, whether a person intends to interact, what attitude they have toward the observer, and which of ten actions they will perform next. To solve it, the authors build SocialEgoNet, a graph-based spatiotemporal network that reads whole-body skeleton keypoints from the face, hands, and body and uses a hierarchical classifier to exploit dependencies among the three tasks. On the augmented JPL-Social dataset, SocialEgoNet attains 83.15% average accuracy across all tasks with real-time inference, outperforming R3D-18 and ST-GCN on every task while being smaller and faster. The work also provides person-level annotations that turn the JPL egocentric interaction dataset into a benchmark for joint social forecasting.

What carries the argument

The central machinery is SocialEgoNet, a three-stage pipeline: (1) AlphaPose extracts whole-body keypoints (68 face, 23 body, 42 hand) and constructs a pose graph for each part; (2) three three-layer GCNs produce spatial embeddings that are concatenated, passed through multi-head self-attention, and fed into a three-layer Bi-LSTM to capture temporal dynamics across the 30-frame window; (3) a hierarchical classifier with a chain structure first predicts intent, then attitude, then action, mimicking the order in which a person sizes up an approaching human. The JPL-Social dataset—the other supporting artifact—adds person-level tracks, bounding boxes, and new labels for gaze and no-response actions to the original JPL videos, yielding 290 annotated person tracks.

What would settle it

Remove the attitude head from SocialEgoNet and, on the held-out test split, map the predicted action labels to attitude labels through the same annotation rule; if the resulting accuracy is statistically indistinguishable from the reported 91.11%, then the attitude task adds no predictive information beyond action classification. A complementary test: have independent annotators label attitude from facial expression and body language without seeing the action label, and check whether SocialEgoNet trained on those labels performs above chance.

Watch

Extended reading notes

Core claim

The central claim is that a lightweight graph network operating purely on whole-body keypoints can jointly forecast a person's intent to interact (interacting, interested, not interested), their attitude (positive or negative), and their upcoming action (ten classes) from the first second of an egocentric interaction, and that doing so jointly with a chain-structured hierarchical classifier outperforms treating the tasks independently. SocialEgoNet achieves 88.10% intent accuracy, 91.11% attitude accuracy, and 70.24% action accuracy, for an average of 83.15%, with 3.18 million parameters and sub-millisecond graph inference speed, beating R3D-18 and ST-GCN; the larger MS-G3D obtains higher accuracy (85.08%) but is about eight times slower. The paper further claims that each body part contributes complementary information: face keypoints dominate intent and attitude prediction, while body and hand keypoints matter most for action prediction.

Load-bearing premise

The attitude label is defined by the action class: punching and throwing count as negative, every other action as positive, so attitude forecasting collapses into action recognition and the high attitude accuracy is not evidence of an independent social signal.

Editorial extensions

If this is right

  • If SocialEgoNet's results hold, robots and embodied agents can use one-second skeleton input to prepare a response before a human reaches them, without using raw RGB frames, preserving privacy and saving compute.
  • The chain hierarchy shows a measurable gain over parallel or single-task heads (82.33% vs 80.13% and 78.93% average F1), supporting the design principle that intent and attitude are useful superclasses for action forecasting.
  • Because face keypoints carry most information for intent and attitude, downstream systems could focus camera resolution on the face region when only those two tasks matter.
  • Opening the observation window from 8 to 30 frames improves attitude and action accuracy more than intent accuracy, suggesting that early intent is readable sooner than valence or specific action.

Reading between the lines

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

  • Since attitude labels in JPL-Social are defined entirely by the action class (punching and throwing negative, everything else positive), the reported 91.11% attitude accuracy should be read as measuring action discrimination; a genuinely independent attitude signal would need human-annotated valence labels not derived from the action taxonomy.
  • The same architecture could be extended to multi-person scenes by running the person-track module on all detected tracks and adding a ranking head that predicts which person is most likely to initiate interaction; this would test whether the intent-to-interact superclass idea scales to social selection.
  • Because the model uses only keypoint coordinates and confidence scores, its transfer to new environments is likely limited by pose-estimation quality on unseen camera angles, occlusions, and unusual clothing; a stress test with noisy or absent keypoints would bound its robustness.
  • The hierarchy assumption that intent is a superclass of attitude and action could be tested causally by intervening on the training labels (shuffling attitude labels) and measuring whether the action head's accuracy drops; that would reveal how much the chain structure truly carries information.
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

4 major / 6 minor

Summary. The paper introduces SocialEgoNet, a graph-based spatiotemporal model that jointly forecasts a person's intent to interact, attitude, and action from one second of egocentric whole-body skeleton video. The authors also introduce JPL-Social, an augmented version of the JPL-Interaction dataset with person-level labels for these three tasks. The model uses separate GCNs for body, face, and hand keypoints, fuses them with multi-head self-attention, and models temporal dependencies with a Bi-LSTM before passing features to a hierarchical classifier. On 290 person tracks, the model achieves an average accuracy of 83.15%, outperforming R3D-18 and ST-GCN but not MS-G3D, while being smaller and faster. The paper reports ablations on body-part combinations, hierarchy designs, and observation window size.

Significance. If the three tasks were genuinely independent social signals, the paper would present a practical real-time framework and a useful dataset contribution. The strengths of the work are the whole-body keypoint design, the real-time inference claim, the body-part and hierarchy ablations, and the planned release of annotations and code. However, the construct validity of the attitude and intent labels is a central issue: attitude is defined directly from the action taxonomy, and the interested/not-interested distinction for non-interacting persons is essentially a gaze/no-response label. As a result, the reported accuracy numbers overstate the degree to which the model forecasts three distinct social cues. The empirical comparisons are plausible but are supported by a single run with no error bars or significance tests.

major comments (4)
  1. [Section III, attitude annotation] The attitude ground truth is a deterministic function of the action class: 'Punching or Throwing an object to the observer are labelled as having a negative attitude while the rest are considered as having a positive attitude.' Therefore the 91.11% attitude accuracy in Table II does not measure an independent social attitude; it measures the model's ability to separate two action classes (punching and throwing) from the other eight. The paper should either collect independent attitude annotations (e.g., from human judges) or explicitly reframe attitude as a label derived from the action taxonomy rather than a separate social signal.
  2. [Section III, intent annotation] The intent-to-interact labels for non-interacting persons are also nearly deterministic: a person is 'interested' when constantly looking at the observer, and the new action class 'gaze' is defined as looking at the observer, while 'no response' is defined as ignoring the observer. Thus the interested/not-interested split is essentially a relabeling of the gaze/no-response action classes. The paper should clarify the exact mapping between intent classes and action classes and discuss the implication that intent, attitude, and action are not three independent annotation dimensions.
  3. [Section V, Table II] The experimental protocol is under-specified: the paper does not state the train/test split, whether the split is at the video level or the person-track level, the number of runs, or any variance statistics. With only 290 person-level samples and strong data augmentation, the reported margins over ST-GCN (2.33% average accuracy) and the hierarchy gains in Table IV may not be statistically reliable. The authors should report repeated runs with means and standard deviations, or at minimum a confidence interval, and describe the split protocol precisely.
  4. [Section IV-C and Table IV] The Chain classifier explicitly encodes the annotation rule that attitude is the parent class of actions, because attitude is defined by whether the action is punching/throwing or not. The observed improvement of Chain over Parallel and Tree may therefore simply reflect the deterministic group structure of the labels, not a learned semantic hierarchy. To support the claim that the hierarchical structure learns complementary information between tasks, the paper needs a control condition (e.g., a hierarchy based on a random grouping of actions) or independent attitude labels.
minor comments (6)
  1. [Section V-A and Table II] The text says the model outperforms 'several competitive baselines,' but MS-G3D achieves the best accuracy on every task. The paper should acknowledge this explicitly and frame the contribution as a favorable accuracy/speed/parameter trade-off rather than overall state-of-the-art accuracy.
  2. [Table II caption and Section V-A] The latency comparison is not fully apples-to-apples: graph-based methods include pose extraction time (Delta), while R3D-18 does not. The paper should state the timing comparison conditions more clearly, including whether pose extraction is included for all methods.
  3. [Section III, 'annotated with three labels each'] The text says every person track is annotated with three labels each, but later states that attitude is applicable only to interacting subjects. This is contradictory; the paper should clarify how attitude is assigned and evaluated for non-interacting tracks.
  4. [Fig. 3] The class distribution figure is difficult to read in the text; providing the exact per-class counts in a table or in the caption would improve reproducibility and transparency.
  5. [Algorithm 1, line 5] The dimension notation 'z(t) = Concat(...) in R^{n x 16}' is unclear because n is not defined; please clarify the shape of the concatenated feature.
  6. [Throughout] The name 'MS-G3D Net' in Section V-A and 'MS-G3D' in Table II should be made consistent.

Circularity Check

2 steps flagged · score 6.0 of 10

Attitude and intent labels are deterministic regroupings of the action taxonomy in Section III, so two of the three 'forecasts' reduce to action discrimination by construction.

  1. self definitional [Section III (JPL-Social: Augmented JPL Dataset), 'Additional Annotations']
    "We create another set of labels that indicates the attitude of a subject towards the observer. This is done based on the actions they perform and applicable to only the 'interacting' subjects. Punching or Throwing an object to the observer are labelled as having a negative attitude while the rest are considered as having a positive attitude (see Fig. 3)."

    By this annotation rule, attitude is a deterministic two-class partition of the action taxonomy: negative exactly equals {punching, throwing}, and positive equals every other action. The attitude head is therefore predicting a derived binary grouping of the action labels, not an independent social attitude. The reported 91.11% attitude accuracy in Table II is the accuracy of that derived binary classifier. The Chain classifier's premise that 'attitude can be regarded as the parent class of some actions' (Section IV-C) merely re-encodes this label construction in the network, so its performance gain is an expected consequence of the annotation scheme, not evidence of a learned dependency between two distinct signals.

  2. self definitional [Section III (JPL-Social: Augmented JPL Dataset), 'Additional Annotations']
    "The video-level labels provided originally are assigned to the interacting person directly. For the subjects not interacting, their tracks are annotated with two new actions: gaze (i.e., looking at the observer) and no response (i.e., focusing on their own business and ignoring the observer). The interested label is assigned to subjects that are not yet interacting but are interested in interacting in the near future. A person is considered interested when constantly looking at the observer (camera), otherwise, the person is labelled as not interested."

    Because 'gaze' is one of the action classes and 'no response' is the other new action class, the intent labels for non-interacting subjects are exactly 'interested = gaze' and 'not interested = no response'; interacting subjects receive one of the eight original action labels. Hence the three-class intent task is a deterministic regrouping of the 10-class action task. The 88.10% intent accuracy in Table II therefore measures action-group discrimination, and averaging it with action accuracy in the 83.15% headline double-counts the same underlying label information.

full rationale

SocialEgoNet is evaluated on held-out person tracks against standard baselines, and the hierarchical-classifier comparisons are empirical, so there is no statistical fitting of the test labels and no load-bearing self-citation chain. The circularity lies in the task construction. Section III defines attitude as a binary function of the action label: negative = {punching, throwing}, positive = every other action. It also defines intent as a three-way grouping of the action labels: the original eight actions map to 'interacting', 'gaze' maps to 'interested', and 'no response' maps to 'not interested'. Thus the attitude and intent forecasts are, by construction, relabelings of the action forecast, and the reported 91.11% attitude accuracy and 88.10% intent accuracy measure the model's ability to separate those action groups rather than independent social signals. The Chain classifier (Section IV-C) encodes the same nesting by treating attitude as the parent class of actions, so the multitask gains in Table IV are expected from the label hierarchy. The average accuracy of 83.15% therefore double-counts the action-discrimination signal across three nominally distinct tasks. This is a self-definitional reduction of two of the three claimed predictions, not a statistical overfitting of a parameter.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on a small lab dataset (200 videos, 290 tracks) whose new labels are partly defined in terms of each other: attitude is derived from action class, and interested intent is defined by gaze. The model uses standard components with hand-chosen hyperparameters; no new physical or mathematical entities are introduced.

free parameters (4)
  • Number of GCN layers per body part = 3
    Hand-chosen; no ablation over depth reported.
  • Number of Bi-LSTM layers = 3
    Hand-chosen; no ablation over depth reported.
  • Observation window length = 30 frames (1 second)
    Fixed by task definition; performance increases with longer windows, so this is a latency-driven choice.
  • Data augmentation ranges = cropping scales [0.95, 0.85, 0.75]; keypoint noise sigma in [0.005, 0.01]
    Hand-chosen; no sensitivity analysis.
assumptions (5)
  • domain assumption AlphaPose pre-trained on COCO2017 yields accurate whole-body keypoints on 320x240 egocentric videos.
    The entire model input depends on this; no pose quality analysis is reported. Section IV-A.
  • domain assumption The JPL lab-recorded interactions are representative of real human-agent interaction.
    Claims of suitability for real-world deployment rest on this; the paper notes future work in-the-wild. Sections III and VI.
  • ad hoc to paper Attitude ground truth can be derived from action classes (punching and throwing negative, all others positive).
    This defines the attitude task and creates a deterministic dependency between attitude and action labels. Section III.
  • ad hoc to paper Interested intent can be operationalized as constantly looking at the observer.
    This equates intent with gaze, making the intent task largely a gaze classification problem. Section III.
  • domain assumption The first second of video contains sufficient information to forecast future actions for all ten classes.
    Observation window is fixed at 30 frames; the paper shows performance increases with longer windows, indicating the one-second window is a limiting choice. Section V-D.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interact with me: Joint Egocentric Forecasting of Intent to Interact, Attitude and Social Actions." pith.science (2026). https://pith.science/paper/ADSC5OGL

@misc{pith2026241216698,
  author       = {Pith},
  title        = {Pith review of: Interact with me: Joint Egocentric Forecasting of Intent to Interact, Attitude and Social Actions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ADSC5OGL}},
  note         = {Machine review of arXiv:2412.16698}
}
read the original abstract

For efficient human-agent interaction, an agent should proactively recognize their target user and prepare for upcoming interactions. We formulate this challenging problem as the novel task of jointly forecasting a person's intent to interact with the agent, their attitude towards the agent and the action they will perform, from the agent's (egocentric) perspective. So we propose \emph{SocialEgoNet} - a graph-based spatiotemporal framework that exploits task dependencies through a hierarchical multitask learning approach. SocialEgoNet uses whole-body skeletons (keypoints from face, hands and body) extracted from only 1 second of video input for high inference speed. For evaluation, we augment an existing egocentric human-agent interaction dataset with new class labels and bounding box annotations. Extensive experiments on this augmented dataset, named JPL-Social, demonstrate \emph{real-time} inference and superior performance (average accuracy across all tasks: 83.15\%) of our model outperforming several competitive baselines. The additional annotations and code will be available upon acceptance.

Figures

Figures reproduced from arXiv: 2412.16698 by the authors.

Figure 1
Figure 1. To understand social interactions for human-agent interactions, we propose the task of jointly forecasting user’s [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed model, SocialEgoNet, for the joint forecasting task we introduce. We first utilize AlphaPose [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. JPL-Social class distribution for the three tasks—intent (in bold), attitude (in italic), and actions (in regular). The number of samples in each class is indicated in parentheses, and newly introduced labels are underscored. person i.e., person-level as opposed to video-level (see Table I and [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The proposed designs of our hierarchical classifiers. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Effect of the observation window size on SocialE [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 31 canonical work pages

  1. [1]

    A survey on the application trends of home service robotics,

    Georgios A Zachiotis, George Andrikopoulos, and George Nikolakopou- los, “A survey on the application trends of home service robotics,” in 2018 IEEE international conference on Robotics and Biomimetics (ROBIO). IEEE, 2018, pp. 1999–2006. 1

    work page 2018

  2. [2]

    Break the ice: a survey on socially aware engagement for human–robot first encounters,

    Jo ˜ao Avelino, Leonel Garcia-Marques, and Alexandre Bernardino, “Break the ice: a survey on socially aware engagement for human–robot first encounters,” International Journal of Social Robotics , vol. 13, no. 8, pp. 1851–1877, 2021. 1

    work page 2021

  3. [3]

    Hierarchical classifier with overlapping class groups,

    Igor T Podolak, “Hierarchical classifier with overlapping class groups,” Expert Systems with Applications , vol. 34, no. 1, pp. 673–682, 2008. 1

    work page 2008

  4. [4]

    First-person activity recognition: What are they doing to me?,

    Michael S Ryoo and Larry Matthies, “First-person activity recognition: What are they doing to me?,” in Proc CVPR, 2013, pp. 2730–2737. 1, 2, 3

    work page 2013

  5. [5]

    Robot-centric activity prediction from first-person videos: What will they do to me?,

    Michael S Ryoo, Thomas J Fuchs, and Larry Matthies, “Robot-centric activity prediction from first-person videos: What will they do to me?,” in Proc ICRA. IEEE, 2015, pp. 295–302. 1, 2, 3

    work page 2015

  6. [6]

    A closer look at spatiotemporal convolutions for action recognition,

    Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri, “A closer look at spatiotemporal convolutions for action recognition,” in Proc CVPR, 2018, pp. 6450–6459. 2, 4, 5

    work page 2018

  7. [7]

    Spatial temporal graph convolutional networks for skeleton-based action recognition,

    Sijie Yan, Yuanjun Xiong, and Dahua Lin, “Spatial temporal graph convolutional networks for skeleton-based action recognition,” in Proc AAAI, 2018. 2, 4, 5

    work page 2018

  8. [8]

    Human action recognition and prediction: A survey,

    Yu Kong and Yun Fu, “Human action recognition and prediction: A survey,” International Journal of Computer Vision , vol. 130, no. 5, pp. 1366–1401, 2022. 2

    work page 2022

Show all 33 references
  1. [9]

    How robots influence humans: A survey of nonverbal communication in social human–robot interaction,

    Shane Saunderson and Goldie Nejat, “How robots influence humans: A survey of nonverbal communication in social human–robot interaction,” International Journal of Social Robotics , vol. 11, no. 4, pp. 575–608,

  2. [10]

    Intention estimation from gaze and motion features for human-robot shared-control object ma- nipulation,

    Anna Belardinelli and Tomoki Watabe, “Intention estimation from gaze and motion features for human-robot shared-control object ma- nipulation,” in Proc International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 9806–9813. 2

    2022

  3. [11]

    Developing an engagement-aware system for the detection of unfocused interaction,

    Marvin Brenner, Heike Brock, Andreas Stiegler, and Randy Gomez, “Developing an engagement-aware system for the detection of unfocused interaction,” in Proc International Conference on Robot & Human Interactive Communication (RO-MAN) . IEEE, 2021, pp. 798–805. 2

    2021

  4. [12]

    Mindread- ing for robots: Predicting intentions via dynamical clustering of human postures,

    Samuele Vinanzi, Christian Goerick, and Angelo Cangelosi, “Mindread- ing for robots: Predicting intentions via dynamical clustering of human postures,” in Proc International conference on development and learning and epigenetic robotics (ICDL-EpiRob) . IEEE, 2019, pp. 272–277. 2

    2019

  5. [13]

    Self- supervised prediction of the intention to interact with a service robot,

    Gabriele Abbate, Alessandro Giusti, Antonio Paolillo, et al., “Self- supervised prediction of the intention to interact with a service robot,” Robotics and Autonomous Systems , vol. 171, pp. 104568, 2024. 2

    2024

  6. [14]

    Systematic review of emotion detection with computer vision and deep learning,

    Rafael Pereira, Carla Mendes, Ant ´onio Pereira, et al., “Systematic review of emotion detection with computer vision and deep learning,” Sensors, vol. 24, no. 11, pp. 3484, 2024. 2

    2024

  7. [15]

    Reducing viral transmission through ai-based crowd moni- toring and social distancing analysis,

    Benjamin Fraser, Brendan Copp, Gurpreet Singh, Orhan Keyvan, Tongfei Bian, Valentin Sonntag, Yang Xing, Weisi Guo, and Antonios Tsourdos, “Reducing viral transmission through ai-based crowd moni- toring and social distancing analysis,” in 2022 IEEE International Con- ference o...

    2022

  8. [16]

    Grassmannian learning mutual subspace method for image set recognition,

    Lincon Souza, Naoya Sogi, Bernardo B Gatto, Takumi Kobayashi, and Kazuhiro Fukui, “Grassmannian learning mutual subspace method for image set recognition,” Neurocomputing, vol. 517, pp. 20–33, 2023. 2

    2023

  9. [17]

    Deep emotion recognition through upper body movements and facial expression.,

    Chaudhary Muhammad Aqdus Ilyas, Rita Nunes, and Thomas B Moes- lund, “Deep emotion recognition through upper body movements and facial expression.,” in Proc VISIGRAPP (5: VISAPP) . INSTICC, 2021, pp. 669–679. 2

    2021

  10. [18]

    Scaling egocentric vision: The epic-kitchens dataset,

    Dima Damen, Hazel Doughty, Will Price, et al., “Scaling egocentric vision: The epic-kitchens dataset,” in Proc ECCV. Springer, 2018, pp. 720–736. 2

    2018

  11. [19]

    Ego4d: Around the world in 3,000 hours of egocentric video,

    Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al., “Ego4d: Around the world in 3,000 hours of egocentric video,” in Proc CVPR, 2022, pp. 18995–19012. 2

    2022

  12. [20]

    Nape: Numbering as a position encoding in graphs,

    Olayinka Ajayi, Hongkai Wen, and Tanaya Guha, “Nape: Numbering as a position encoding in graphs,” IEEE Access , vol. 12, pp. 166200– 166210, 2024. 2

    2024

  13. [21]

    Rolling-unrolling lstms for action anticipation from first-person video,

    Antonino Furnari and Giovanni Maria Farinella, “Rolling-unrolling lstms for action anticipation from first-person video,” IEEE transactions on pattern analysis and machine intelligence , vol. 43, no. 11, pp. 4021– 4036, 2020. 2

    2020

  14. [22]

    Predicting the next action by modeling the abstract goal,

    Debaditya Roy and Basura Fernando, “Predicting the next action by modeling the abstract goal,” in Proc ICPR. 2024, p. 162–177, Springer- Verlag. 2

    2024

  15. [23]

    Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition,

    Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer, “Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition,” in Proc CVPR, 2022, pp. 13587–13597. 2

    2022

  16. [24]

    AntGPT: Can large language models help long-term action anticipation from videos?,

    Qi Zhao, Shijie Wang, Ce Zhang, Changcheng Fu, Minh Quan Do, Nakul Agarwal, Kwonjoon Lee, and Chen Sun, “AntGPT: Can large language models help long-term action anticipation from videos?,” in Proc ICLR,

  17. [25]

    Can’t make an omelette without breaking some eggs: Plausible action anticipation using large video-language models,

    Himangi Mittal, Nakul Agarwal, Shao-Yuan Lo, and Kwonjoon Lee, “Can’t make an omelette without breaking some eggs: Plausible action anticipation using large video-language models,” in Proc CVPR, 2024, pp. 18580–18590. 2

    2024

  18. [26]

    Yolov3: An incremental improve- ment,

    Joseph Redmon and Ali Farhadi, “Yolov3: An incremental improve- ment,” arXiv preprint arXiv:1804.02767 , 2018. 3

    2018 arXiv

  19. [27]

    Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time,

    Hao-Shu Fang, Jiefeng Li, Cewu Lu, et al., “Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2022. 4, 5

    2022

  20. [28]

    Microsoft coco: Common objects in context,

    Tsung-Yi Lin, Michael Maire, and C Lawrence Zitnick, “Microsoft coco: Common objects in context,” in Proc ECCV. Springer, 2014, pp. 740–755. 4

    2014

  21. [29]

    Semi-supervised classification with graph convolutional networks,

    Thomas N. Kipf and Max Welling, “Semi-supervised classification with graph convolutional networks,” in Proc ICLR, 2017. 4

    2017

  22. [30]

    Attention is all you need,

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin, “Attention is all you need,” in Proc NeurIPS . 2017, p. 6000–6010, Curran Associates Inc. 4

    2017

  23. [31]

    Framewise phoneme classifi- cation with bidirectional lstm networks,

    Alex Graves and J ¨urgen Schmidhuber, “Framewise phoneme classifi- cation with bidirectional lstm networks,” in Proc International Joint Conference on Neural Networks , 2005, vol. 4, pp. 2047–2052. 4

    2005

  24. [32]

    Dg-stgcn: Dy- namic spatial-temporal modeling for skeleton-based action recognition,

    Haodong Duan, Jiaqi Wang, Kai Chen, and Dahua Lin, “Dg-stgcn: Dy- namic spatial-temporal modeling for skeleton-based action recognition,” arXiv preprint arXiv:2210.05895 , 2022. 4, 5

    2022 arXiv

  25. [33]

    Disentangling and unifying graph convolutions for skeleton- based action recognition,

    Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang, “Disentangling and unifying graph convolutions for skeleton- based action recognition,” in Proc CVPR, 2020. 4, 5

    2020

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.