OSGNet @ Ego4D Episodic Memory Challenge 2025

Haoyu Zhang; Liqiang Nie; Meng Liu; Qiaohui Chu; Weili Guan; Yaowei Wang; Yisen Feng

REVIEW 3 major objections 4 minor 26 references

OSGNet @ Ego4D Episodic Memory Challenge 2025

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read OSGNet, an early-fusion grounding network that mixes video, text, and detected-object features, claims first place on all three Ego4D Episodic Memory Challenge tracks: Natural Language Queries, Goal Step, and Moment Queries.

desk verdict Credible first-place challenge results, but the paper's early-fusion claim outruns its evidence—fine as an engineering write-up, thin as a scientific claim. read the letter →

arxiv 2506.03710 v1 pith:K2YIASK5 submitted 2025-06-04 cs.CV cs.AI

Yisen Feng , Haoyu Zhang , Qiaohui Chu , Meng Liu , Weili Guan , Yaowei Wang , Liqiang Nie This is my paper

classification cs.CVcs.AI

keywords Ego4Degocentricvideomomentlocalizationtemporalgroundingnaturallanguagequeriesgoalstepearlyfusionobject-enhancednetwork

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

OSGNet is a single unified video-localization architecture entered in all three Ego4D Episodic Memory Challenge tracks—Natural Language Queries, Goal Step, and Moment Queries—and the paper reports that it took first place in each. The explanation offered is early fusion: object-level cues from a detector are combined with video and text features through attention and gated fusion before the localization heads, rather than being merged late as in earlier unified systems. On the test sets it reports gains over last year's winners of 2.47 Rank@1 points on NLQ at IoU=0.5, 6.84 Rank@1 points on Goal Step at IoU=0.3, and 1.79 mAP points on Moment Queries. Further gains come from ensembling three or four model variants and adding the validation set to training. If the attribution to early fusion is right, object-aware early fusion is a better basis for egocentric moment localization than the late-fusion unified baselines it compares against.

What carries the argument

The load-bearing object is the OSGNet grounding network, used here with its early-fusion design. Its main branch fuses three modalities—video, query text, and object detections—through object attention, cross attention, and gated fusion before a multi-scale temporal network and regression/classification heads; its shot branch pools video-only features into shot segments and applies contrastive learning with positive and negative shots to strengthen cross-modal alignment. The challenge recipe wraps this core with NaQ narration-as-query pretraining, EgoVideo or InternVideo features, random cropping for very long Goal Step videos, a Bayesian temporal-order prior at inference, and ensembles of three to four model variants. Early fusion of object features is the mechanism the paper explicitly credits for its accuracy.

What would settle it

Train OSGNet and a late-fusion variant on the Ego4D validation split using the same EgoVideo and CLIP features, the same object features, the same training data, and no ensemble or validation-set augmentation, and compare Rank@1 and mAP; if the late-fusion variant matches or beats OSGNet, the early-fusion explanation is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that early fusion, reinforced with object-level detail, is what allows one model to localize moments accurately across three egocentric-video task formulations. Concretely, OSGNet extracts video, text, and detected-object features (using CLIP, EgoVideo or InternVideo, and Co-DETR), then fuses them with object attention, cross attention, and gate fusion inside a multi-scale network before classification and regression heads predict the temporal interval; a separate shot branch aligns video-only shot features to text through contrastive learning. The reported test results place OSGNet first on all three tracks: Rank@1 at IoU=0.5 of 21.78 versus 19.31 for EgoVideo on NLQ, Rank@1 at IoU=0.3 of 42.02 versus 35.18 for BayesianVSLNet on Goal Step, and 36.78 versus 34.99 mAP for CausalTAD on Moment Queries, with ensembles and validation-set training included in the winning numbers.

Load-bearing premise

The load-bearing premise is that the reported lead comes from early fusion itself, but no matched ablation against a late-fusion baseline is run, so competing explanations—ensemble, validation-set training, and stronger video features—are not ruled out.

Editorial extensions

If this is right

If the reported results hold, a single early-fusion model can lead all three Ego4D localization tracks without task-specific localization heads.
Fine-grained object information becomes a first-order cue: models that ignore object detections should trail on Ego4D-style egocentric localization.
The reported margins over last year's champions set a new reference level that future entries to these tracks will need to beat.
Ensembling variants and adding the validation set to training adds 1.73 Rank@1 points on NLQ, 1.51 on Goal Step, and 1.47 mAP points on MQ, so the single-model numbers understate the deployed system's performance.
Reformulating action localization as text-based retrieval costs extra training and inference time, a bottleneck the paper itself flags in its conclusion.

Reading between the lines

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

The paper leaves untested whether early fusion itself causes the win; a matched comparison against a late-fusion baseline with identical features and data would be needed to separate architecture from ensemble and training-set effects.
Its own failure cases suggest the practical ceiling is the object detector's vocabulary and the model's composition of subject-object relations, so better detector coverage and relational reasoning should translate directly into higher localization scores.
The same object-enhanced early-fusion recipe should transfer to exocentric long-video grounding and video question answering, which share the need to bind fine-grained objects to temporal intervals; the paper does not test this.
Object class scores may also serve as weak spatial and temporal anchors for attribute- or location-style queries, an interpretation the paper leaves implicit.

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.

Desk Editor's Note

read the letter

First: the first-place claims are credible and the write-up is honest about what it is. Second: the paper's one scientific sentence—that early fusion 'demonstrates its effectiveness'—is not supported by the evidence in the tables. The stress-test note lands; I'd go further and say the paper reads exactly like a strong challenge report, not a research claim.

What's actually new: OSGNet itself is from the authors' arXiv:2505.04270, so the architecture is reused. The new pieces are the task translations—GoalStep and MQ turned into text-based moment localization—and the specific ensemble/validation-set recipe that produced the winning numbers. Those numbers are externally checkable against the Ego4D leaderboard, and the code is linked. The paper also does the right thing by naming its known limitations: object-detector coverage failures and the training/inference cost of retrieval reformulation.

Where it's soft: the central attribution. The best rows in Tables 1-3 are ensembles of up to four variants, trained on validation data, with multiple feature backbones. The single-model rows already beat last year's champions, but those baselines differ in pretraining, features, and inference priors. There is no controlled comparison that swaps only the fusion stage. So 'early fusion wins' is a hypothesis, not a demonstrated result. The margins (2.47, 6.84, 1.79) are plausible but come without error bars or significance tests. The paper's own limitation paragraph mentions object coverage and cost but never flags the attribution problem, so the abstract overclaims by presenting first place as proof of the architecture's superiority.

None of this is fatal to the report's purpose. For someone in egocentric video localization, it's a useful record of what worked at the 2025 challenge, and the implementation details are specific enough to reproduce. The reader's conditional verdict seems right to me; I wouldn't ask the authors to add a full late-fusion ablation if this stays a challenge report, but if it's submitted as a regular paper, that ablation is the minimum.

Recommendation: engage with it as a challenge write-up. It deserves a referee if venues treat competition reports as citable empirical records; desk-rejecting it would lose the concrete SOTA numbers. Just don't let 'first place' be read as evidence for early fusion without a controlled experiment.

Referee Report

3 major / 4 minor

Summary. This challenge report describes OSGNet, an early-fusion egocentric video localization model applied to three Ego4D Episodic Memory Challenge 2025 tracks: Natural Language Queries (NLQ), GoalStep-Step Grounding, and Moment Queries (MQ). The authors report first-place results in all three tracks, with gains over the previous champions of 2.47% R@1@0.5 on NLQ, 6.84% R@1@0.3 on GoalStep, and 1.79% mAP on MQ. The system uses video, text, and object features from EgoVideo, InternVideo, and CLIP, with a main branch and a shot branch, following the authors' prior OSGNet work. The best reported results are obtained by ensembling multiple model variants and including the validation set in training. The paper's central claim is that early fusion (as opposed to late fusion) is responsible for the performance.

Significance. If the first-place claims are accurate, this is a strong and externally verifiable result on a high-profile public benchmark. The paper also ships code and provides per-track implementation details (features, training schedules, ensemble composition), which aids reproducibility. However, the scientific significance beyond the challenge results is limited by the absence of any controlled comparison isolating the early-fusion mechanism. The gains are attributed to the architecture, but the experimental design conflates fusion strategy with feature backbones, pretraining, ensemble size, and training set composition. The paper is honest about some limitations (object detector coverage, retrieval re-reformulation cost) but does not acknowledge that the reported margins do not support the early-fusion attribution without an ablation.

major comments (3)

[Abstract and Section 2.2, Tables 1-3] The central claim that OSGNet's early-fusion design 'demonstrates its effectiveness' is not supported by the reported experiments. The best numbered rows in Tables 1-3 are 'OSGNet†', which are ensembles of up to four variants (Section 2.2) trained with the validation set. The reported margins over EgoVideo, BayesianVSLNet, and CausalTAD therefore conflate the fusion strategy with multiple simultaneous changes: video features (EgoVideo/InternVideo), NaQ pretraining, object annotations, training epochs, ensemble size, and validation-set inclusion. A controlled ablation comparing OSGNet with a late-fusion variant under matched features, pretraining, and training data is required before any causal role can be attributed to early fusion. Without such an experiment, the phrase 'demonstrating its effectiveness' overclaims.
[Section 2.3, Table 2] The interpretation of the GoalStep results is muddled by the unstated interaction with the BayesianVSLNet order prior. The text says 'During inference, we incorporate the order prior as introduced in BayesianVSLNet,' but it is not clarified whether the validation numbers in Table 2 (44.27 R@1@0.3) include this prior. Since the comparison baseline also uses the prior, the reported 6.84% improvement could partly arise from how the prior interacts with OSGNet's predictions. The paper should state explicitly whether the prior is applied in each reported row, or provide numbers with and without it.
[Table 2 validation vs. test] The validation and test columns for BayesianVSLNet in Table 2 appear internally inconsistent: the validation R@1@0.3 is 18.15 while the test value is 35.18, and the validation R@1@0.5 is 8.97 while the test value is 20.48. The corresponding OSGNet values are roughly comparable between validation and test (44.27 vs. 40.51 and 37.37 vs. 31.55). This discrepancy suggests either a typo, a different validation partition, or a different evaluation protocol for the baseline. Please clarify, because the GoalStep comparison is a load-bearing evidence for the claimed improvement.

minor comments (4)

[Section 2.2, Implementation Details] The sentence 'we first train OSGNet without the shot branch and object features (named OSGNet-baseline)' is confusing because the next sentence says the pretraining uses NaQ, and then the main training phase is described. It would be clearer to specify exactly which components are frozen or reinitialized between pretraining and fine-tuning.
[Section 2.4, Table 3] The footnote for Table 3 defines R1@0.5 as 'R@1, IoU=0.5 metric,' but this notation is used inconsistently; the abstract and Section 2.2 use 'Rank@1 at IoU=0.5.' Please standardize the metric naming.
[Section 2.2, Figure 2 and Section 2.3, Figure 3] The success/failure examples are informative, but the figure captions only state 'Two examples on the validation set' without describing what the reader should observe. One or two sentences per example explaining the temporal alignment (or misalignment) with the ground truth would make the figures self-contained.
[References] Reference [19] duplicates [18] with a slightly different page format; please merge or remove one. Also, reference [4] is cited for both the OSGNet architecture and the claimed state-of-the-art results; if that paper is under review, a note on availability would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported first-place results are externally benchmarked, and the self-citations to OSGNet are disclosed references to an independently evaluated model, not definitions of the reported metrics.

full rationale

The paper's central claim is empirical: OSGNet and its ensembled variants achieved first place on the Ego4D Episodic Memory Challenge tracks. These results are compared against external challenge baselines and evaluated on a held-out test set, so they are not derived from the model's own equations or fitted to the target metric. The method section refers to the authors' prior OSGNet paper for architectural details, and this self-citation is load-bearing only in the sense of identifying the model used; it does not define the challenge metrics or make the leaderboard outcome equivalent to the architecture by construction. No step in the report reduces a prediction to a fitted parameter, renames a known result, or imports a uniqueness theorem. The claim that early fusion is the cause of the improvement is under-supported because no matched late-fusion ablation is presented and the gains conflate ensemble, validation-set training, and stronger features, but that is a correctness/attribution concern, not circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the availability and quality of pretrained models (EgoVideo, InternVideo, CLIP, Co-DETR) and on the Ego4D benchmark evaluations. No new mathematical or physical axioms are introduced; the free parameters are training hyperparameters and design choices.

free parameters (4)

learning rates = 8e-4 pretrain, 1.6e-3 main NLQ/GoalStep, 4e-4 MQ
Training hyperparameters chosen by the authors, not derived from theory.
batch sizes = 16 for NLQ, 32 for GoalStep and MQ
Training setup choices.
training epochs = 10 with 4 warm-up epochs
Training duration, set manually.
max frame features for GoalStep = 2560
Cap applied during random cropping to manage computational cost.

assumptions (3)

domain assumption Pretrained feature extractors (EgoVideo, InternVideo, CLIP) encode sufficient temporal and semantic information for egocentric moment localization.
The paper uses these features as input to OSGNet without evaluating their contribution ablationally.
domain assumption The Co-DETR object detector can detect the objects referenced in natural language queries and action descriptions.
The paper notes that detection failure for untrained categories ('fire gun') causes localization failure, confirming that this assumption is load-bearing.
ad hoc to paper The order prior from BayesianVSLNet improves GoalStep inference without harming other tracks.
The order prior is applied at test time for GoalStep only, based on the prior method's success, not on a derivation in this report.

how reviews work

0 comments

Cite this review

Pith. "Pith review of OSGNet @ Ego4D Episodic Memory Challenge 2025." pith.science (2026). https://pith.science/paper/K2YIASK5

@misc{pith2026250603710,
  author       = {Pith},
  title        = {Pith review of: OSGNet @ Ego4D Episodic Memory Challenge 2025},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K2YIASK5}},
  note         = {Machine review of arXiv:2506.03710}
}

read the original abstract

In this report, we present our champion solutions for the three egocentric video localization tracks of the Ego4D Episodic Memory Challenge at CVPR 2025. All tracks require precise localization of the interval within an untrimmed egocentric video. Previous unified video localization approaches often rely on late fusion strategies, which tend to yield suboptimal results. To address this, we adopt an early fusion-based video localization model to tackle all three tasks, aiming to enhance localization accuracy. Ultimately, our method achieved first place in the Natural Language Queries, Goal Step, and Moment Queries tracks, demonstrating its effectiveness. Our code can be found at https://github.com/Yisen-Feng/OSGNet.

Figures

Figures reproduced from arXiv: 2506.03710 by the authors.

**Figure 1.** The framework of OSGNet [4]. and encoding of features from video, text, and detected objects, thereby enabling more accurate moment localization. The framework consists of two key components: a main branch that fuses video, text, and object features to improve localization precision, and a shot branch that processes video-only features to generate shot-level representations, leveraging contrastive learning with t… view at source ↗

**Figure 2.** Two examples on the validation set of NLQ. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Two examples on the validation set of Goal Step. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Example on the validation set of MQ. the training set, outperforms last year’s winning solution. By incorporating the validation set and applying model ensembling, the performance is further improved by 1.47% in mAP. Specifically, the ensemble includes three model var…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 12 canonical work pages

[1]

Internvideo-ego4d: A pack of champion solu- tions to ego4d challenges

Guo Chen, Sen Xing, Zhe Chen, Yi Wang, Kunchang Li, Yizhuo Li, Yi Liu, Jiahao Wang, Yin-Dong Zheng, Bingkun Huang, et al. Internvideo-ego4d: A pack of champion solu- tions to ego4d challenges. arXiv preprint arXiv:2211.09529, pages 1–11, 2022. 2

arXiv 2022
[2]

Technical report for ego4d long-term action anticipation challenge 2025

Qiaohui Chu, Haoyu Zhang, Yisen Feng, Meng Liu, Weili Guan, Yaowei Wang, and Liqiang Nie. Technical report for ego4d long-term action anticipation challenge 2025. arXiv preprint arXiv:2506.02550, 2025. 1

arXiv 2025
[3]

ObjectNLQ@ Ego4D episodic mem- ory challenge 2024

Yisen Feng, Haoyu Zhang, Yuquan Xie, Zaijing Li, Meng Liu, and Liqiang Nie. ObjectNLQ@ Ego4D episodic mem- ory challenge 2024. arXiv preprint arXiv:2406.15778, 2024. 1

arXiv 2024
[4]

Object-shot enhanced grounding network for egocentric video

Yisen Feng, Haoyu Zhang, Meng Liu, Weili Guan, and Liqiang Nie. Object-shot enhanced grounding network for egocentric video. arXiv preprint arXiv:2505.04270, 2025. 1, 2, 3

arXiv 2025
[5]

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. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995–19012, 2022. 1

2022
[6]

Bi-directional het- erogeneous graph hashing towards efficient outfit recom- mendation

Weili Guan, Xuemeng Song, Haoyu Zhang, Meng Liu, Chung-Hsing Yeh, and Xiaojun Chang. Bi-directional het- erogeneous graph hashing towards efficient outfit recom- mendation. In Proceedings of the 30th ACM international conference on multimedia, pages 268–276, 2022. 1

work page 2022
[7]

Groundnlq@ ego4d natural language queries challenge 2023

Zhijian Hou, Lei Ji, Difei Gao, Wanjun Zhong, Kun Yan, Chao Li, Wing-Kwong Chan, Chong-Wah Ngo, Nan Duan, and Mike Zheng Shou. Groundnlq@ ego4d natural language queries challenge 2023. arXiv preprint arXiv:2306.15255 , pages 1–5, 2023. 1

arXiv 2023
[8]

Attentive moment retrieval in videos

Meng Liu, Xiang Wang, Liqiang Nie, Xiangnan He, Bao- quan Chen, and Tat-Seng Chua. Attentive moment retrieval in videos. In The 41st international ACM SIGIR conference on research & development in information retrieval , pages 15–24, 2018. 1

work page 2018

Show all 26 references

[9]

Cross-modal moment localiza- tion in videos

Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localiza- tion in videos. In Proceedings of the 26th ACM international conference on Multimedia, pages 843–851, 2018. 1

2018
[10]

Harnessing temporal causal- ity for advanced temporal action detection

Shuming Liu, Lin Sui, Chen-Lin Zhang, Fangzhou Mu, Chen Zhao, and Bernard Ghanem. Harnessing temporal causal- ity for advanced temporal action detection. arXiv preprint arXiv:2407.17792, 2024. 1

2024 arXiv
[11]

EgoVideo: Exploring egocentric founda- tion model and downstream adaptation

Baoqi Pei, Guo Chen, Jilan Xu, Yuping He, Yicheng Liu, Kanghua Pan, Yifei Huang, Yali Wang, Tong Lu, Limin Wang, et al. EgoVideo: Exploring egocentric founda- tion model and downstream adaptation. arXiv preprint arXiv:2406.18070, pages 1–8, 2024. 1, 2, 3

2024 arXiv
[12]

CARLOR@ Ego4D step grounding challenge: Bayesian temporal-order priors for test time re- finement

Carlos Plou, Lorenzo Mur-Labadia, Ruben Martinez-Cantin, and Ana C Murillo. CARLOR@ Ego4D step grounding challenge: Bayesian temporal-order priors for test time re- finement. arXiv preprint arXiv:2406.09575 , pages 1–4,

arXiv
[13]

NaQ: Leveraging narrations as queries to su- pervise episodic memory

Santhosh Kumar Ramakrishnan, Ziad Al-Halah, and Kris- ten Grauman. NaQ: Leveraging narrations as queries to su- pervise episodic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 6694–6703. IEEE Computer Society, 2023. 2

2023
[14]

Ego4d goal-step: To- ward hierarchical understanding of procedural activities.Ad- vances in Neural Information Processing Systems, 36, 2024

Yale Song, Eugene Byrne, Tushar Nagarajan, Huiyu Wang, Miguel Martin, and Lorenzo Torresani. Ego4d goal-step: To- ward hierarchical understanding of procedural activities.Ad- vances in Neural Information Processing Systems, 36, 2024. 1

2024
[15]

Time: Temporal-sensitive multi-dimensional instruc- tion tuning and benchmarking for video-llms

Yunxiao Wang, Meng Liu, Rui Shao, Haoyu Zhang, Bin Wen, Fan Yang, Tingting Gao, Di Zhang, and Liqiang Nie. Time: Temporal-sensitive multi-dimensional instruc- tion tuning and benchmarking for video-llms. arXiv preprint arXiv:2503.09994, 2025. 1

2025 arXiv
[16]

Unimd: Towards unifying moment retrieval and temporal ac- tion detection

Yingsen Zeng, Yujie Zhong, Chengjian Feng, and Lin Ma. Unimd: Towards unifying moment retrieval and temporal ac- tion detection. In European Conference on Computer Vision, pages 286–304. Springer, 2025. 1

2025
[17]

TimeLoc: A uni- fied end-to-end framework for precise timestamp localiza- tion in long videos

Chen-Lin Zhang, Lin Sui, Shuming Liu, Fangzhou Mu, Zhangcheng Wang, and Bernard Ghanem. TimeLoc: A uni- fied end-to-end framework for precise timestamp localiza- tion in long videos. arXiv preprint arXiv:2503.06526, 2025. 1

2025 arXiv
[18]

Multimodal dialog system: Rela- tional graph-based context-aware question understanding

Haoyu Zhang, Meng Liu, Zan Gao, Xiaoqiang Lei, Yinglong Wang, and Liqiang Nie. Multimodal dialog system: Rela- tional graph-based context-aware question understanding. In Proceedings of the 29th ACM International Conference on Multimedia, page 695–703. Association for Computin...

2021
[19]

Multimodal dialog system: Rela- tional graph-based context-aware question understanding

Haoyu Zhang, Meng Liu, Zan Gao, Xiaoqiang Lei, Yinglong Wang, and Liqiang Nie. Multimodal dialog system: Rela- tional graph-based context-aware question understanding. In Proceedings of the 29th ACM international conference on multimedia, pages 695–703, 2021

2021
[20]

Attribute-guided collab- orative learning for partial person re-identification

Haoyu Zhang, Meng Liu, Yuhong Li, Ming Yan, Zan Gao, Xiaojun Chang, and Liqiang Nie. Attribute-guided collab- orative learning for partial person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14144–14160, 2023

2023
[21]

Uncovering hidden connections: Iterative tracking and reasoning for video-grounded dialog

Haoyu Zhang, Meng Liu, Yaowei Wang, Da Cao, Weili Guan, and Liqiang Nie. Uncovering hidden connections: Iterative tracking and reasoning for video-grounded dialog. arXiv preprint arXiv:2310.07259, 2023. 1

2023 arXiv
[22]

Multi-factor adaptive vision selec- tion for egocentric video question answering

Haoyu Zhang, Meng Liu, Zixin Liu, Xuemeng Song, Yaowei Wang, and Liqiang Nie. Multi-factor adaptive vision selec- tion for egocentric video question answering. InProceedings of the 41st International Conference on Machine Learning , pages 59310–59328. PMLR, 2024. 1

2024
[23]

Hcqa@ ego4d egoschema challenge

Haoyu Zhang, Yuquan Xie, Yisen Feng, Zaijing Li, Meng Liu, and Liqiang Nie. Hcqa@ ego4d egoschema challenge
[24]

Exo2ego: Exocentric knowledge guided mllm for egocentric video understanding

Haoyu Zhang, Qiaohui Chu, Meng Liu, Yunxiao Wang, Bin Wen, Fan Yang, Tingting Gao, Di Zhang, Yaowei Wang, and Liqiang Nie. Exo2ego: Exocentric knowledge guided mllm for egocentric video understanding. arXiv preprint arXiv:2503.09143, 2025. 1

2025
[25]

Hcqa- 1.5@ ego4d egoschema challenge 2025

Haoyu Zhang, Yisen Feng, Qiaohui Chu, Meng Liu, Weili Guan, Yaowei Wang, and Liqiang Nie. Hcqa- 1.5@ ego4d egoschema challenge 2025. arXiv preprint arXiv:2505.20644, 2025. 1

2025 arXiv
[2024]

arXiv preprint arXiv:2406.15771, 2024. 1

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Internvideo-ego4d: A pack of champion solu- tions to ego4d challenges

Guo Chen, Sen Xing, Zhe Chen, Yi Wang, Kunchang Li, Yizhuo Li, Yi Liu, Jiahao Wang, Yin-Dong Zheng, Bingkun Huang, et al. Internvideo-ego4d: A pack of champion solu- tions to ego4d challenges. arXiv preprint arXiv:2211.09529, pages 1–11, 2022. 2

arXiv 2022

[2] [2]

Technical report for ego4d long-term action anticipation challenge 2025

Qiaohui Chu, Haoyu Zhang, Yisen Feng, Meng Liu, Weili Guan, Yaowei Wang, and Liqiang Nie. Technical report for ego4d long-term action anticipation challenge 2025. arXiv preprint arXiv:2506.02550, 2025. 1

arXiv 2025

[3] [3]

ObjectNLQ@ Ego4D episodic mem- ory challenge 2024

Yisen Feng, Haoyu Zhang, Yuquan Xie, Zaijing Li, Meng Liu, and Liqiang Nie. ObjectNLQ@ Ego4D episodic mem- ory challenge 2024. arXiv preprint arXiv:2406.15778, 2024. 1

arXiv 2024

[4] [4]

Object-shot enhanced grounding network for egocentric video

Yisen Feng, Haoyu Zhang, Meng Liu, Weili Guan, and Liqiang Nie. Object-shot enhanced grounding network for egocentric video. arXiv preprint arXiv:2505.04270, 2025. 1, 2, 3

arXiv 2025

[5] [5]

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. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995–19012, 2022. 1

2022

[6] [6]

Bi-directional het- erogeneous graph hashing towards efficient outfit recom- mendation

Weili Guan, Xuemeng Song, Haoyu Zhang, Meng Liu, Chung-Hsing Yeh, and Xiaojun Chang. Bi-directional het- erogeneous graph hashing towards efficient outfit recom- mendation. In Proceedings of the 30th ACM international conference on multimedia, pages 268–276, 2022. 1

work page 2022

[7] [7]

Groundnlq@ ego4d natural language queries challenge 2023

Zhijian Hou, Lei Ji, Difei Gao, Wanjun Zhong, Kun Yan, Chao Li, Wing-Kwong Chan, Chong-Wah Ngo, Nan Duan, and Mike Zheng Shou. Groundnlq@ ego4d natural language queries challenge 2023. arXiv preprint arXiv:2306.15255 , pages 1–5, 2023. 1

arXiv 2023

[8] [8]

Attentive moment retrieval in videos

Meng Liu, Xiang Wang, Liqiang Nie, Xiangnan He, Bao- quan Chen, and Tat-Seng Chua. Attentive moment retrieval in videos. In The 41st international ACM SIGIR conference on research & development in information retrieval , pages 15–24, 2018. 1

work page 2018