A Probabilistic Jump-Diffusion Framework for Open-World Egocentric Activity Recognition

Sanjoy Kundu; Sathyanarayanan N. Aakur; Shanmukha Vellamcheti

REVIEW 4 major objections 4 minor 11 references

A Probabilistic Jump-Diffusion Framework for Open-World Egocentric Activity Recognition

T0 review · 4 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 Open-world egocentric activity recognition can be done with an order of magnitude fewer VLM queries by replacing brute-force label scans with a prior-guided jump-diffusion search.

desk verdict The L1 efficiency claim and the L0-L3 taxonomy are worth taking seriously; the L2/L3 numbers are not checkable until the Gemini search spaces are released. read the letter →

arxiv 2505.22858 v1 pith:UE7BYF3G submitted 2025-05-28 cs.CV

Sanjoy Kundu , Shanmukha Vellamcheti , Sathyanarayanan N. Aakur This is my paper

classification cs.CV

keywords egocentricactivityrecognitionopen-worldvision-languagemodelsjump-diffusionsearchcommonsenseknowledgepriorsConceptNetVLMqueryefficiencytaxonomy

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

The paper claims that open-world egocentric activity recognition can be done accurately without scanning every possible label, by replacing brute-force VLM enumeration with a stochastic jump-diffusion search over a commonsense-structured label space. The proposed ProbRes framework builds a candidate search space from ConceptNet priors over action-object pairs, then alternates prior-guided exploration with likelihood-driven exploitation and a final action/object re-ranking. Across four egocentric benchmarks (GTEA Gaze, GTEA Gaze+, EPIC-Kitchens-100, Charades-Ego) and openness levels L1-L3, ProbRes reports higher or comparable accuracy than exhaustive VLM querying while cutting VLM calls from thousands to hundreds. The upshot is that scalable open-world recognition depends on search strategy, not on enumerating an unbounded label space, and the paper offers a four-level taxonomy (L0-L3) for saying how open a task is.

What carries the argument

The load-bearing mechanism is a jump-diffusion search over the VLM text-embedding space $\varphi(a)$; 'jump-diffusion' means each step alternates large exploratory jumps drawn from the ConceptNet-derived prior with local likelihood-driven moves, so the trajectory escapes poor regions and still converges on high-likelihood labels. The search space is pre-structured by ConceptNet semantic-affinity scores and sorted by embedding distance to an anchor, giving the jumps semantic locality. A local refinement step deterministically re-evaluates top candidates, and a concept-decomposition step scores action and object components separately with $v^\top \varphi(\cdot)$; the final ranking combines video-text likelihood with component scores via $S_{\text{final}} = P_{\text{likelihood}}(v|a) + \lambda_a S_a + \lambda_o S_o$. This is what lets the search cover tens to hundreds of thousands of label candidates in a few hundred VLM calls.

What would settle it

Regenerate the L3 search space on EPIC-Kitchens with the ground-truth activity labels removed from the candidate list, then run ProbRes; if accuracy and WUPS collapse, the candidate list's coverage of the true labels, rather than the search mechanism, is driving the reported results.

Watch

Extended reading notes

Core claim

The central claim is that the argmax of a VLM likelihood over an unconstrained activity space—the formal target of open-world recognition—can be reached by jump-diffusion rather than enumeration. ProbRes derives a prior $P_{\text{prior}}(a)$ over action-object pairs from ConceptNet semantic affinity, organizes the text-embedding space $\varphi(a)$ around an anchor, and iterates between prior-guided exploration samples and likelihood-weighted exploitation samples before re-ranking survivors with $S_{\text{final}} = P_{\text{likelihood}}(v|a) + \lambda_a S_a + \lambda_o S_o$. The empirical result is that this procedure improves activity phrase accuracy and WUPS over exhaustive VLM baselines on GTEA Gaze, GTEA Gaze+, EPIC-Kitchens-100, and Charades-Ego, while using as few as 110 VLM queries where the baseline used 380 and 3,000 where the baseline used 29,100. The paper also establishes a four-level taxonomy of openness (L0-L3), arguing that different search-space constraints require different methodological responses.

Load-bearing premise

The L2 and L3 results assume that the Gemini 2.0 Flash-generated candidate lists contain the true activity labels; because neither the lists nor the generation prompts are released, missing labels or dataset-tuned lists would make the reported gains disappear in a genuinely open world.

Editorial extensions

If this is right

Open-world egocentric recognition no longer requires enumerating every candidate label, because a VLM queried only on prior-guided, likelihood-refined candidates can beat exhaustive scanning.
VLM query cost drops by an order of magnitude or more at every openness level, from 380 to 110 queries on GTEA Gaze at L1 up to 195,714 to 5,000 on EPIC-Kitchens at L3.
The L0-L3 taxonomy gives researchers a shared scale for how open a recognition setting is, separating fixed label sets, compositional novelty, domain-specific open search, and fully open search.
Fixed commonsense knowledge graphs alone are insufficient for open-world settings; the priors must be combined with likelihood-driven adaptation to handle dynamic, domain-specific spaces.
VLM text embeddings are not well organized enough for direct nearest-neighbor search, so open-world systems need the extra refinement and re-ranking step that ProbRes adds.

Reading between the lines

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

The same exploration-exploitation recipe should transfer to other open-vocabulary tasks with a semantic label graph and a VLM likelihood, such as open-world object detection or compositional human-object interaction classification.
ProbRes's query counts stay roughly constant as the search space grows from hundreds to hundreds of thousands of candidates, so its relative advantage should be largest exactly where exhaustive enumeration is most impractical.
Publishing the Gemini-generated L2/L3 candidate lists and prompts would convert the reported results into a reproducible benchmark and let outsiders test whether coverage of the lists, rather than the search itself, explains the accuracy.
The reported two-second inference on an RTX 3090 suggests near-real-time use is plausible, but deployment on wearable or embedded egocentric devices would still need its own latency and energy measurements, which the paper does not provide.

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

Two things to know before reading this one. First, the L1 result is a plausible engineering contribution: on four egocentric datasets, probing a fixed action-object space with ConceptNet priors and VLM likelihoods cuts VLM calls 3-10x while slightly improving accuracy. Second, the L2/L3 headline claim rests on Gemini 2.0 Flash-generated candidate spaces that are never described, released, or contamination-checked, so those numbers should be treated as unverified until the artifacts appear.

What's actually new: the L0-L3 openness taxonomy is useful framing for egocentric recognition, and the specific combination of a structured prior, iterative exploration/exploitation, and VLM likelihood scoring is not in the prior work I know. The efficiency-accuracy tradeoff matters for real-time robotics and wearable AI, and if it holds, it is a practical step forward.

The paper also does some things well. The prior comes from ConceptNet and the likelihood from pretrained VLMs, so the core result is not forced by the paper's own normalization. The limitations section is honest: it says LLM-generated search spaces 'require careful curation,' which is precisely the weak point. The L1 tables include the strong exhaustive baselines and show improvements that are plausible if not airtight.

Soft spots. The L2/L3 evaluation is the load-bearing weakness. The entire open-world claim is that ProbRes finds the true label inside a space the authors generated with Gemini. The paper reports no prompts, no candidate lists, no analysis of coverage, no contamination check. If a generated space omitted the true label, ProbRes could never retrieve it; if the space was curated with knowledge of the ground truth, the comparison against exhaustive LA VILA/EGOVLP enumeration is not a fair open-world test. The stress-test note is right: as reported, the L2/L3 efficiency-accuracy tradeoff is unfalsifiable. In addition, the search parameters (lambda, T, lambda_a, lambda_o) are tuned for efficiency and accuracy without a stated held-out protocol, so the reported tradeoff could be a selected optimum rather than a general property. Finally, the jump-diffusion framing is a misnomer—the equations are weighted sampling with a refinement step; no jump or diffusion process appears. That's a naming issue, not a fatal flaw, but it should be fixed.

Who this is for: researchers working on open-world egocentric recognition and anyone who wants to reduce VLM query costs. The taxonomy and the L1 evidence justify sending this to a serious referee. I'd ask for the Gemini prompts, candidate lists, and ground-truth coverage analysis, plus a clear tuning protocol, before accepting the results.

Recommendation: accept for peer review, with the expectation of a substantive revision that makes the L2/L3 artifacts public.

Referee Report

4 major / 4 minor

Summary. The paper introduces ProbRes, a probabilistic residual search framework for open-world egocentric activity recognition. The method combines ConceptNet-based prior probabilities with VLM likelihoods in an iterative exploration-exploitation search, followed by local refinement and re-ranking, to recognize activities without exhaustively enumerating all possible labels. The authors define a four-level openness taxonomy (L0–L3), evaluate on GTEA Gaze, GTEA Gaze+, EPIC-Kitchens-100, and Charades-Ego, and report accuracy, WUPS, and VLM-call counts. For L1, ProbRes is shown to reduce VLM calls while improving or maintaining accuracy relative to exhaustive LA VILA and EGOVLP inference. For L2 and L3, search spaces are generated by Gemini 2.0 Flash and the method again reports large reductions in VLM calls with accuracy gains.

Significance. If the results hold, the paper makes a useful contribution: it demonstrates that structured commonsense priors can guide VLM-based search in open-world recognition, and it introduces a taxonomy that could help standardize evaluations. The L1 experiments provide reasonably direct evidence that the approach can reduce query counts while preserving accuracy, and the use of ConceptNet as an external prior avoids the circularity concerns that sometimes plague self-referential open-world benchmarks. However, the L2/L3 results currently rest on unstated generation protocols and unreleased candidate spaces, and all quantitative comparisons lack error bars; these issues must be resolved before the headline efficiency-accuracy claim can be accepted.

major comments (4)

[Sec. 3.2, Implementation Details; Sec. 6; Table 2] The L2/L3 evaluation is not reproducible as reported. The paper states that search spaces are 'generated using Gemini 2.0 Flash' but gives no prompts, no candidate lists, no release of the generated spaces, and no protocol for ensuring that the generator did not see or was not tuned to the ground-truth labels. The Limitations section itself says LLM-generated search spaces 'require careful curation,' yet no curation or contamination check is described. Since ProbRes can only retrieve labels present in the candidate space, the reported WUPS and accuracy for L2/L3 are ceilinged by the generator. Please provide the generation prompts, example candidate lists, a coverage analysis showing how often true labels appear in the generated spaces, and a statement of whether any filtering or regeneration was done with knowledge of the test labels.
[Sec. 3.2, Implementation Details; Tables 1 and 2] The hyperparameters (lambda, T, lambda_a, lambda_o) are tuned for both efficiency and accuracy, but no held-out protocol is described. This is load-bearing because T directly determines the reported VLM-call counts: if T is selected per dataset to maximize accuracy under a budget, the 'VLM calls' numbers become a tuned output rather than a property of the method. Please specify the validation split, the tuning grid, and the selection criterion, and report sensitivity of accuracy to T. Alternatively, fix T as a budget and do not tune it per dataset.
[Tables 1 and 2] All reported numbers appear to come from a single run with no error bars, no seeds, and no statistical significance tests. Several of the claimed gains are small (e.g., Table 2, EGOVLP+ProbRes on GTEA Gaze L3 activity: 36.47 vs. 35.94; and Table 1, EGOVLP+ProbRes on EK100 activity: 40.53 vs. 39.94), so without variance estimates it is unclear whether these differences are meaningful. Please report mean and standard deviation over at least three seeds, or justify why the comparisons are deterministic.
[Sec. 3.2, Eq. (2) and Eq. (3)] The core optimization objective is not fully defined. Equation (2) uses Psearch(a), but the text only defines Pexplore(a) and Pguided(a); it is never stated whether Psearch is one of these, a sum, or a phase-dependent quantity. Equation (3) also contains an unclear notation: the denominator appears to be '1/|S| P a' ...' with the P symbol serving both as a probability and as a summation index. Please define Psearch explicitly and rewrite Eq. (3) with unambiguous summation limits.

minor comments (4)

[Table 2] The row '59.2530.3644.81' appears to lack spaces between three numbers; it should read '59.25 30.36 44.81'.
[Table 1] The label 'LA VILA-Decomp380' and 'EGOVLP-Decomp380' should include a space before the number (e.g., 'LA VILA-Decomp 380') for consistency with other entries.
[Figure 2 caption] The caption contains a typo: 'ProbRes frameworkfor' should be 'ProbRes framework for'.
[Sec. 3.2, Implementation Details] The statement 'Average inference time is 2 seconds per video on an RTX 3090' would be more useful with clarification of whether this includes Gemini search-space generation and VLM calls, and over which dataset and openness level it was measured.

Circularity Check

0 steps flagged · score 1.0 of 10

No demonstrated circularity: ProbRes's derivation is self-contained (external ConceptNet priors and pretrained-VLM likelihoods), self-citations are baselines only, and the final re-ranking interpolates the same VLM's scores rather than defining the output as its input.

full rationale

Walking the claimed derivation chain: the optimization target (Eq. 1) is the VLM likelihood, and the search objective (Eq. 2) combines a prior derived from ConceptNet (an external knowledge base, Ref. [8]) with that likelihood and with action/object component scores S_a and S_o computed as v^T phi(.) from the same VLM backbone. The final score S_final = P_likelihood(v|a) + lambda_a S_a + lambda_o S_o is a re-ranking or interpolation of the same backbone's scores, not an identity: the components are computed on decomposed (action, object) text, and the reported gains over exhaustive VLM enumeration indicate that the re-weighting changes rankings. No parameter is defined in terms of the ground-truth label, and no uniqueness theorem or ansatz is imported from the authors' prior work; the mechanism cites external sources (ConceptNet [8], EGOVLP [6], LA VILA [10]). The self-citations (Refs. [1] and [4], both by overlapping authors) appear only as experimental baselines (KGL, ALGO) and are not load-bearing for the ProbRes mechanism. Per the review rule, the appended limitation in Sec. 6 is weighed explicitly: 'LLM-generated search spaces, while enabling scalability, require careful curation.' This is directly relevant to the L2/L3 claims, because the L2/L3 search space is 'generated using Gemini 2.0 Flash' (Sec. 3.2) and Eq. 1 restricts the argmax to that space S, so the reported accuracy is structurally ceilinged by list membership. However, the prompts and candidate lists are not released, so no specific reduction (e.g., ground-truth leakage into the candidate lists) is exhibited; per the hard rules, this is a verification and contamination risk (a correctness concern), not a demonstrated circular step. Similarly, tuning (lambda, T, lambda_a, lambda_o) on the target data is a fitting concern, not a definitional circularity. Verdict: no significant circularity; the derivation is self-contained against external benchmarks and prior resources.

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

The central claims rest on external knowledge resources and on generated candidate spaces. The ConceptNet prior and VLM likelihood are not derived in the paper; they are treated as given. The L2/L3 experiments further assume that Gemini-generated candidate lists adequately represent the true open-world label space. These are domain assumptions that the paper does not validate, and their failure would change the reported results.

free parameters (4)

lambda (exploration-exploitation weight) = not reported, tuned per dataset
Controls the mixture of the ConceptNet prior and uniform sampling in Eqn 3, directly shaping the search trajectory.
T (search budget, number of VLM calls) = 110-5,000 depending on dataset and openness level
The headline VLM-call reductions are a direct consequence of choosing T; the paper tunes T for efficiency and accuracy, so the reported call counts are partly selected rather than predicted.
lambda_a (action component weight) = not reported, tuned per dataset
Weight on the action-alignment score S_a in Eqn 2 and the final ranking; tuned per dataset.
lambda_o (object component weight) = not reported, tuned per dataset
Weight on the object-alignment score S_o in Eqn 2 and the final ranking; tuned per dataset.

assumptions (3)

domain assumption ConceptNet semantic affinity f(a_action, a_object) provides a useful prior P_prior(a) over activities.
Sec. 3.1; if the prior is poorly aligned with the video domains, exploration is misdirected and the search budget could be wasted.
domain assumption VLM text embedding similarity v^T phi(a) is a reliable likelihood for ranking candidate activities.
Eqs. 1-4; the whole search exploits this likelihood, and the paper itself notes VLM embedding inconsistencies in Sec. 6.
ad hoc to paper Gemini-generated candidate search spaces for L2/L3 include or semantically cover the true ground-truth activity labels.
Sec. 3.2 and Tables 2; no candidate lists or prompts are released, and the open-world accuracy results depend entirely on this coverage assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Probabilistic Jump-Diffusion Framework for Open-World Egocentric Activity Recognition." pith.science (2026). https://pith.science/paper/UE7BYF3G

@misc{pith2026250522858,
  author       = {Pith},
  title        = {Pith review of: A Probabilistic Jump-Diffusion Framework for Open-World Egocentric Activity Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UE7BYF3G}},
  note         = {Machine review of arXiv:2505.22858}
}

read the original abstract

Open-world egocentric activity recognition poses a fundamental challenge due to its unconstrained nature, requiring models to infer unseen activities from an expansive, partially observed search space. We introduce ProbRes, a Probabilistic Residual search framework based on jump-diffusion that efficiently navigates this space by balancing prior-guided exploration with likelihood-driven exploitation. Our approach integrates structured commonsense priors to construct a semantically coherent search space, adaptively refines predictions using Vision-Language Models (VLMs) and employs a stochastic search mechanism to locate high-likelihood activity labels while minimizing exhaustive enumeration efficiently. We systematically evaluate ProbRes across multiple openness levels (L0--L3), demonstrating its adaptability to increasing search space complexity. In addition to achieving state-of-the-art performance on benchmark datasets (GTEA Gaze, GTEA Gaze+, EPIC-Kitchens, and Charades-Ego), we establish a clear taxonomy for open-world recognition, delineating the challenges and methodological advancements necessary for egocentric activity understanding.

Figures

Figures reproduced from arXiv: 2505.22858 by the authors.

**Figure 1.** Taxonomy of Openness in Egocentric Activity Recognition. We define four levels of openness based on search space constraints: L0 (fixed activity set), L1 (known atomic concepts, unknown compositions), L2 (known domain, inferred activities), and L3 (fully unconstrained search). structured priors with VLM-based likelihood refinement to efficiently navigate unconstrained activity spaces, drastically reducing VLM quer… view at source ↗

**Figure 2.** (a) ProbRes framework for open-world egocentric activity recognition. The search space is structured using ConceptNet priors, enabling guided exploration. The model iteratively refines candidates via likelihood estimation, balancing exploration and exploitation, followed by local refinement and re-ranking. (b) Search trajectory visualization, showing how ProbRes navigates likelihood regions to reach high-confidence … view at source ↗

**Figure 3.** Qualitative Visualization of the search trajectory by ProbRes across different phases indicating exploration, exploitation, and the final refinement phase. to 110 on GTEA Gaze and 29,100 to 3,000 on EK100, while improving WUPS and phrase accuracy. This efficiencyaccuracy trade-off is vital for scalability, demonstrating effective navigation without brute-force enumeration, even on larger datasets. Neuro-symbolic … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

11 extracted references · 8 canonical work pages

[1]

Aakur, Sanjoy Kundu, and Nikhil Gunti

Sathyanarayanan N. Aakur, Sanjoy Kundu, and Nikhil Gunti. Knowledge guided learning: Open world egocentric action recognition with zero supervision.Pattern Recognition Let- ters, 156:38–45, 2022. Publisher: North-Holland. 3

work page 2022
[2]

The EPIC-KITCHENS Dataset: Collection, Challenges and Baselines, 2020

Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. The EPIC-KITCHENS Dataset: Collection, Challenges and Baselines, 2020. arXiv:2005.00343 [cs]. 3

arXiv 2020
[3]

Alireza Fathi, Yin Li, and James M. Rehg. Learning to Recognize Daily Actions Using Gaze. InComputer Vision – ECCV 2012, pages 314–327, Berlin, Heidelberg, 2012. Springer. 3

work page 2012
[4]

Sanjoy Kundu, Shubham Trehan, and Sathyanarayanan N. Aakur. Discovering Novel Actions from Open World Ego- centric Videos with Object-Grounded Visual Commonsense Reasoning, 2024. arXiv:2305.16602 [cs]. 3

work page Pith review arXiv 2024
[5]

Yin Li, Alireza Fathi, and James M. Rehg. Learning to Pre- dict Gaze in Egocentric Video. In2013 IEEE International Conference on Computer Vision, pages 3216–3223, Sydney, Australia, 2013. IEEE. 3

work page 2013
[6]

Egocentric Video-Language Pretraining,

Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, Chengfei Cai, Hongfa Wang, Dima Damen, Bernard Ghanem, Wei Liu, and Mike Zheng Shou. Egocentric Video-Language Pretraining,

work page
[7]

Actor and Observer: Joint Modeling of First and Third-Person Videos

Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Actor and Observer: Joint Modeling of First and Third-Person Videos, 2018. arXiv:1804.09627 [cs]. 3

work page Pith review arXiv 2018
[8]

Concept- Net 5.5: An Open Multilingual Graph of General Knowl- edge.Proceedings of the AAAI Conference on Artificial In- telligence, 31(1), 2017

Robyn Speer, Joshua Chin, and Catherine Havasi. Concept- Net 5.5: An Open Multilingual Graph of General Knowl- edge.Proceedings of the AAAI Conference on Artificial In- telligence, 31(1), 2017. Number: 1. 2

work page 2017

Show all 11 references

[9]

Verb semantics and lexical selection

Zhibiao Wu and Martha Palmer. Verb semantics and lexical selection. InProceedings of the 32nd annual meeting on Association for Computational Linguistics, pages 133–138,
[10]

Learning Video Representations from Large Language Models, 2022

Yue Zhao, Ishan Misra, Philipp Kr ¨ahenb¨uhl, and Rohit Gird- har. Learning Video Representations from Large Language Models, 2022. arXiv:2212.04501 [cs]. 1, 2, 3 4

2022 arXiv
[2022]

arXiv:2206.01670 [cs]. 2, 3

arXiv

Pith tools

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

[1] [1]

Aakur, Sanjoy Kundu, and Nikhil Gunti

Sathyanarayanan N. Aakur, Sanjoy Kundu, and Nikhil Gunti. Knowledge guided learning: Open world egocentric action recognition with zero supervision.Pattern Recognition Let- ters, 156:38–45, 2022. Publisher: North-Holland. 3

work page 2022

[2] [2]

The EPIC-KITCHENS Dataset: Collection, Challenges and Baselines, 2020

Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. The EPIC-KITCHENS Dataset: Collection, Challenges and Baselines, 2020. arXiv:2005.00343 [cs]. 3

arXiv 2020

[3] [3]

Alireza Fathi, Yin Li, and James M. Rehg. Learning to Recognize Daily Actions Using Gaze. InComputer Vision – ECCV 2012, pages 314–327, Berlin, Heidelberg, 2012. Springer. 3

work page 2012

[4] [4]

Sanjoy Kundu, Shubham Trehan, and Sathyanarayanan N. Aakur. Discovering Novel Actions from Open World Ego- centric Videos with Object-Grounded Visual Commonsense Reasoning, 2024. arXiv:2305.16602 [cs]. 3

work page Pith review arXiv 2024

[5] [5]

Yin Li, Alireza Fathi, and James M. Rehg. Learning to Pre- dict Gaze in Egocentric Video. In2013 IEEE International Conference on Computer Vision, pages 3216–3223, Sydney, Australia, 2013. IEEE. 3

work page 2013

[6] [6]

Egocentric Video-Language Pretraining,

Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, Chengfei Cai, Hongfa Wang, Dima Damen, Bernard Ghanem, Wei Liu, and Mike Zheng Shou. Egocentric Video-Language Pretraining,

work page

[7] [7]

Actor and Observer: Joint Modeling of First and Third-Person Videos

Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Actor and Observer: Joint Modeling of First and Third-Person Videos, 2018. arXiv:1804.09627 [cs]. 3

work page Pith review arXiv 2018

[8] [8]

Concept- Net 5.5: An Open Multilingual Graph of General Knowl- edge.Proceedings of the AAAI Conference on Artificial In- telligence, 31(1), 2017

Robyn Speer, Joshua Chin, and Catherine Havasi. Concept- Net 5.5: An Open Multilingual Graph of General Knowl- edge.Proceedings of the AAAI Conference on Artificial In- telligence, 31(1), 2017. Number: 1. 2

work page 2017