Pith. sign in

REVIEW 4 major objections 5 minor 77 references

EgoPrivacy: What Your First-Person Camera Says About You?

T0 review · 4 major / 5 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 This paper shows that first-person video exposes wearer demographics even with no training, and that fine-tuning or matching third-person footage reveals identity, location, and time.

desk verdict A genuinely useful benchmark, but the headline privacy-leakage numbers are inflated by a weak baseline comparison and need re-baselining. read the letter →

arxiv 2506.12258 v1 pith:JNDTZAXZ submitted 2025-06-13 cs.CV cs.CY

Yijiang Li , Genpei Zhang , Jiacheng Cheng , Yi Li , Xiaojun Shan , Dashan Gao , Jiancheng Lyu , Yuan Li
show 2 more authors
Ning Bi Nuno Vasconcelos
This is my paper
classification cs.CVcs.CY
keywords egocentricvideoprivacywearerfirst-personretrieval-augmentedattackdemographicinferenceidentityretrievalsituationalbenchmark
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's central claim is that a first-person camera does not hide its wearer: gender, race, age, identity, location, and recording time leak through hands, scene, and motion even when no face is visible. To measure this, it builds EgoPrivacy, a large paired first-person/third-person benchmark with demographic, identity, scene, and moment labels, and evaluates adversaries with zero, fine-tuning, and retrieval-augmented capabilities. Off-the-shelf foundation models already classify gender at up to 73% and race at up to 65% accuracy with no fine-tuning, and fine-tuned retrieval models reach top-1 hit rates of 81% for wearer identity and 89% for scene. The proposed Retrieval-Augmented Attack (RAA) pulls third-person clips of the target from an external pool and aggregates their demographic predictions with the egocentric one, improving attack accuracy by up to roughly 16 percentage points. The reason to care is that the threat does not require sophisticated data collection: some leakage is available to anyone who can run a public model.

What carries the argument

The load-bearing mechanism is the ego-to-exo retrieval embedding, a joint embedding learned with supervised contrastive loss that maps egocentric and exocentric clips into one space. Positive pairs are defined by the task: same-identity clips for individual privacy and synchronized space-time pairs for situational privacy, which lets one embedding serve both retrieval tasks and the first stage of RAA. In RAA, the adversary uses this embedding to rank an external pool of third-person clips, takes the top-M hits, runs the demographic classifier on each retrieved clip, and aggregates the results with the egocentric prediction by soft voting. The EgoPrivacy benchmark carries the evaluation by supplying demographic, identity, scene, and moment labels plus in-distribution and out-of-distribution splits, and the paper uses its progressive-masking analysis to attribute the leakage largely to the wearer's hands and other visible biometric markers.

What would settle it

Build a retrieval pool that is the same size and domain but deliberately excludes any clip of the target person, then rerun RAA: if the demographic accuracy gain over the egocentric-only baseline disappears, the boost really comes from finding the target; if it persists, the mechanism is demographic clustering and a defense should target attribute grouping rather than identity retrieval.

Watch

Extended reading notes

Core claim

In the paper's terms, egocentric video is a serious wearer-privacy channel because it records what the wearer sees and does, which is saturated with the wearer's biometric, behavioral, and situational traces. The paper defines three privacy types—demographic (gender, race, age), individual (wearer identity), and situational (scene and moment)—instantiated as seven tasks on 5,625 Ego-Exo4D clips and 4,000 Charades-Ego clips. Its headline empirical finding is that zero-shot foundation models recover demographic attributes at far above chance (up to 73% gender, 65% race, 80% age), that fine-tuning lifts ego-to-ego identity retrieval to 81% and scene retrieval to 89% top-1 hit rate, and that the Retrieval-Augmented Attack—retrieve the wearer's third-person clips, then combine predictions—consistently strengthens demographic classification, sometimes beyond the exocentric-only baseline. The benchmark is designed to support in-distribution and out-of-distribution evaluation as well as four escalating adversary capabilities, so the leakage findings are measured under varied assumptions about attacker data.

Load-bearing premise

The advertised gain of the Retrieval-Augmented Attack assumes an attacker has a pool of third-person videos that actually contains the target and resembles the query; if that pool is absent, contaminated, or out-of-distribution, the RAA boost should shrink, while the zero-shot and fine-tuned leakage results would stand on their own.

Editorial extensions

If this is right

  • Wearer privacy cannot be assumed safe merely because faces and full bodies are out of frame; off-the-shelf zero-shot models already recover gender, race, and age at rates far above chance.
  • An adversary with a small labeled set can do substantially more: fine-tuned models reach 81% top-1 wearer identification and 89% scene retrieval, so de-anonymization and location tracking are practical with modest data.
  • Retrieval-augmented attacks show that the privacy risk of egocentric footage is compounded by public third-person footage: if the same person appears in any external camera pool, demographic leakage can be pushed beyond what either view alone provides.
  • Out-of-distribution results indicate that fine-tuned attacks overfit their training source while zero-shot attacks keep their accuracy, meaning the hardest-to-defend attack is the one requiring the least attacker investment.
  • Temporal modeling and more frames help attackers, so defenses must consider motion and hand dynamics, not just static privacy filters on frames.

Reading between the lines

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

  • The RAA deltas in the paper likely depend on the external pool actually containing the target identity and matching the query's distribution; a pool of random or foreign third-person footage would probably shrink the gains, though the zero-shot demographic results would stand.
  • Because the paper shows top-1 retrieval often selects a different person with the same demographic attributes, the aggregation stage may be acting less as identity matching and more as demographic reinforcement; this suggests a targeted defense could perturb demographic clustering rather than identity retrieval.
  • The same retrieve-then-predict pipeline could be extended beyond demographics to identity or scene classification, since the embedding already supports ego-to-exo retrieval for those tasks; the paper does not test that extension.
  • The surprising drop in egocentric age and race accuracy under Capability 4 (knowing which clips belong to one identity) hints that identity-level ensembling is not automatically beneficial; per-clip classifier noise may be amplified when aggregated.
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 / 5 minor

Summary. This paper introduces EgoPrivacy, a benchmark for measuring privacy leakage from egocentric video. It defines three privacy types (demographic, individual, and situational), seven tasks, and four attacker capabilities, and it evaluates zero-shot and fine-tuned foundation models on those tasks. The paper also proposes Retrieval-Augmented Attack (RAA), a two-stage attack that first retrieves exocentric clips of the same wearer and then aggregates demographic predictions from both views. The headline findings are that zero-shot models achieve up to 73% gender, 65% race, and 80% age accuracy; fine-tuned retrieval reaches 81% identity and 89% scene top-1 hit rate; and RAA improves demographic attack accuracy by up to 16 points.

Significance. The benchmark is a useful contribution: it is substantially larger than prior egocentric privacy benchmarks, uses an identity-disjoint train/test split, includes an out-of-distribution test set, and ships code and data. The RAA idea is novel and clearly relevant to real-world threats. The paper also conducts useful ablations on temporal modeling, voting mechanisms, and frame masking. However, the central privacy-leakage claims are currently overstated: the zero-shot demographic results are not compared to the majority-class prior, and the abstract attributes fine-tuned retrieval performance to zero-shot models. These issues need to be fixed before the paper's conclusions can be accepted.

major comments (4)
  1. [Section 6.2, Table 2] The claim that all attack models are 'higher than random chance by a large margin (more than 15%)' is made against random chance (50%, 33.33%) rather than against the majority-class prior that the paper itself reports. On the OOD test set, the prior accuracies are 60.74% (gender), 54.17% (race), and 79.48% (age). The best zero-shot results are 73.15%, 65.36%, and 79.64%, giving margins of only +12.4, +11.2, and +0.16 points; the age result is statistically indistinguishable from always predicting the majority class. The conclusion that 'an adversary can still perform a successful attack with up to 80% success rate' is therefore not supported for age and is much weaker than advertised for gender and race. The paper should report margins over the majority-class baseline and qualify the zero-shot demographic claims accordingly.
  2. [Abstract vs. Table 3] The abstract states that 'foundation models can effectively compromise wearer privacy even in zero-shot settings by recovering attributes such as identity, scene, gender, and race with 70–80% accuracy.' Table 3 shows zero-shot ego-to-ego identity HR@1 of 0.49–4.85% and zero-shot scene HR@1 of 13.60–28.64%, while the fine-tuned values are 81.25% and 89.21% respectively. The 70–80% figures are only plausible for demographic attributes (and only for gender and race, and only against random chance), not for identity or scene. The abstract must be corrected to avoid attributing fine-tuned retrieval numbers to zero-shot performance.
  3. [Section 5.2 and Table 2] RAA's reported improvements (deltas up to 16.08 points in Table 2) depend on the structural assumption that the external pool Dretr contains the target identity and is distributionally similar to the query. The experiments retrieve from such a pool; if the pool does not contain the target, or is contaminated or out-of-distribution, the augmentation gain is expected to shrink or disappear. The paper should report RAA performance as a function of retrieval accuracy (e.g., by varying the fraction of pool entries that are the target identity), and explicitly state this limitation in the main text rather than only in the threat-model definition.
  4. [Tables 2, 3, 5, B.1] All results are reported as single numbers with no error bars, confidence intervals, or repeated-seed runs. For example, several RAA deltas in Table 2 are 0.00–1.00 points, and the Table 5 voting-ablation differences are of similar magnitude. Without variance estimates or significance tests, it is impossible to tell which of the reported improvements are real. The authors should add repeated runs (or at least standard errors) for the main tables.
minor comments (5)
  1. [Section 1] Typo: 'suveilance' should be 'surveillance'.
  2. [Tables 2 and 3] Model names are rendered inconsistently ('VideoMAEB', 'VideoMAEL', 'CLIPH/14'); the table formatting should be cleaned up.
  3. [Section 3.2 / Appendix A] The annotation confidence threshold of 80% is described, but no inter-annotator agreement statistics are reported; adding them would strengthen the demographic labels.
  4. [Section 6.1] Training details mention learning rate, batch size, and optimizer but not the number of training epochs or the random seed; include these for reproducibility.
  5. [Appendix B] Capability 4 results are presented as a drop on egocentric age and race; since this is surprising, a brief explanation or analysis would help.

Circularity Check

1 steps flagged · score 2.0 of 10

Core benchmark measurements are independent; the only localized circularity is test-set selection of RAA voting hyperparameters.

  1. fitted input called prediction [Section 6.2, 'Ablation study on voting parameters' (Table 5 and Figure 3)]
    "We also ablate the effect of the k in top-k retrieval in Figure 3, where k= 3 leads to the optimal performance for Gender and Age. For Race, we observe that a larger k= 3 leads to increasing performance."

    The RAA deltas reported in Table 2 are presented as evidence of RAA's effectiveness, but the voting hyperparameters k and w are chosen by maximizing accuracy on the same OOD test set used to report those deltas (Figure 3 and Table 5). The paper then says 'we adopt soft voting as the default ensemble method.' The headline gains, such as VideoMAE's +16.08 on gender, are selected maxima over a test-set search rather than held-out predictions. Part of the reported RAA advantage is therefore a fitted parameter presented as an attack result. This is localized to the RAA contribution and does not taint the independent zero-shot or fine-tuned benchmark measurements.

full rationale

The core benchmark derivation is self-contained: demographic labels come from MTurk annotations of exocentric videos, the train/test split is identity-disjoint, and the zero-shot demographic numbers are produced by off-the-shelf models, so these measurements do not reduce to their inputs. The RAA contribution has one localized circular step: k and the soft-voting weight w are selected on the same OOD test set whose accuracy is then reported as RAA's gain, so the headline deltas are partly fitted values. The paper's remaining problems are correctness or consistency issues rather than derivation circularity. The abstract's claim of 'identity, scene, gender, and race with 70-80% accuracy' is internally contradicted by Table 3, where zero-shot identity HR@1 is 0.5-5% and scene HR@1 is 14-28%, and the best zero-shot age result (79.64%) barely exceeds the disclosed majority prior (79.48%), meaning some demographic 'leakage' may reflect class imbalance. These issues deserve scrutiny, but they are not cases where an equation or defined quantity reduces to its own input, so they do not raise the circularity score beyond the minor RAA hyperparameter-fitting concern.

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

No new physical or theoretical entities are introduced. The contributions are a dataset/benchmark and an attack procedure, not a new force, particle, dimension, or conserved quantity. The free parameters are the tuning choices for RAA and the annotation quality filter. The axioms are the threat-model assumptions and the trust placed in the source datasets' identity splits.

free parameters (4)
  • Soft-voting weight w for the egocentric prediction in RAA = 0.5 (default; 1/(k+1) also tested)
    Chosen via test-set ablation in Table 5; affects the reported RAA deltas.
  • Top-k retrieval count in RAA = 3 (optimal in Figure 3)
    Selected from test-set ablation; RAA accuracy varies with k.
  • Supervised contrastive temperature tau in Eq. 3 = not reported
    Hyperparameter of the ego-to-exo embedding loss; value is not stated in the paper.
  • MTurk annotation confidence threshold = 80%
    Appendix A filters annotator responses below 80% confidence; affects label quality and therefore measured attack accuracy.
assumptions (4)
  • domain assumption Perceived demographic labels annotated from exocentric video are treated as ground truth for wearer privacy.
    Appendix A: MTurk workers label gender, race, and age from 3-4 exocentric videos per identity. The paper acknowledges these are perceived, not self-reported, attributes.
  • ad hoc to paper The adversary with Capability 3 has access to an external pool of exocentric videos that contains the target identity and matches the query distribution.
    Section 4.1 and Section 5.2; RAA's evaluation retrieves from a pool that includes the target. If such a pool does not exist or is out-of-distribution, the reported gains may not transfer.
  • domain assumption The Ego-Exo4D train/test split is identity-disjoint.
    Section 3.2: each participant is assigned exclusively to one split, preventing identity leakage into training. Trusted from the source dataset.
  • standard math Cosine similarity in a contrastively learned ego-exo embedding space is a valid retrieval signal for identity, scene, and moment.
    Equation 3 uses supervised contrastive learning; retrieval quality is an empirical property, not a mathematical guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EgoPrivacy: What Your First-Person Camera Says About You?." pith.science (2026). https://pith.science/paper/JNDTZAXZ

@misc{pith2026250612258,
  author       = {Pith},
  title        = {Pith review of: EgoPrivacy: What Your First-Person Camera Says About You?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JNDTZAXZ}},
  note         = {Machine review of arXiv:2506.12258}
}
read the original abstract

While the rapid proliferation of wearable cameras has raised significant concerns about egocentric video privacy, prior work has largely overlooked the unique privacy threats posed to the camera wearer. This work investigates the core question: How much privacy information about the camera wearer can be inferred from their first-person view videos? We introduce EgoPrivacy, the first large-scale benchmark for the comprehensive evaluation of privacy risks in egocentric vision. EgoPrivacy covers three types of privacy (demographic, individual, and situational), defining seven tasks that aim to recover private information ranging from fine-grained (e.g., wearer's identity) to coarse-grained (e.g., age group). To further emphasize the privacy threats inherent to egocentric vision, we propose Retrieval-Augmented Attack, a novel attack strategy that leverages ego-to-exo retrieval from an external pool of exocentric videos to boost the effectiveness of demographic privacy attacks. An extensive comparison of the different attacks possible under all threat models is presented, showing that private information of the wearer is highly susceptible to leakage. For instance, our findings indicate that foundation models can effectively compromise wearer privacy even in zero-shot settings by recovering attributes such as identity, scene, gender, and race with 70-80% accuracy. Our code and data are available at https://github.com/williamium3000/ego-privacy.

Figures

Figures reproduced from arXiv: 2506.12258 by the authors.

Figure 1
Figure 1. Overview of the proposed EgoPrivacy benchmark. What can you tell about the camera wearer from egocentric videos alone? It may come as a surprise that a fair amount of information about the user, such as demographics, identity, time and location of recording, can be inferred from their first-person view footages, despite not revealing their faces or full body. camera is not framing its user. Given the limited attenti… view at source ↗
Figure 2
Figure 2. Retrieval-Augmented Privacy Attacks. In general, several exocentric samples are associated with a single egocentric sample, either because the exocentric video is collected from multiple viewpoints or by definition of the retrieval task. For example, in individual privacy attacks all exocentric videos of the same camera wearer are considered successful retrievals, independently of whether they were shot at the same … view at source ↗
Figure 3
Figure 3. Performance of Retrieval Augmented Attack versus k. upper bound of an egocentric privacy attack. We also notice that, even when the exocentric performance is lower than egocentric, RAA still offers improvements in some cases. We derive a hypothesis that RAA does not need the retrieval model to select the correct identity necessary to improve, but rather the retrieval model will cluster and group identities of simila… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: (1) Increasing the number of frames improves performance (4 ⇒ 8), but saturates beyond 8 or 16 frames; (2) Temporal modeling (Attention or RNN) consistently outperforms MLP. This effect is more pronounced. These findings are further validated for Identity and Situation…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

77 extracted references · 67 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  3. [3]

    11k hands: gender recognition and biometric identification using a large dataset of hand images

    Afifi, M. 11k hands: gender recognition and biometric identification using a large dataset of hand images. Multimedia Tools and Applications, 78 0 (15): 0 20835--20854, 2019

    work page 2019

  4. [4]

    and Borji, A

    Ardeshir, S. and Borji, A. Egocentric meets top-view. IEEE transactions on pattern analysis and machine intelligence (TPAMI), 41 0 (6): 0 1353--1366, 2018 a

    work page 2018

  5. [5]

    and Borji, A

    Ardeshir, S. and Borji, A. Integrating egocentric videos in top-view surveillance videos: Joint identification and temporal alignment. In ECCV, 2018 b

    work page 2018

  6. [6]

    S., and Rauterberg, M

    Betancourt, A., Morerio, P., Regazzoni, C. S., and Rauterberg, M. The evolution of first person vision methods: A survey. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 25 0 (5): 0 744--760, 2015

    work page 2015

  7. [7]

    Bitouk, D., Kumar, N., Dhillon, S., Belhumeur, P., and Nayar, S. K. F ace S wapping: A utomatically R eplacing F aces in P hotographs. ACM Transactions on Graphics (ToG), 2008

    work page 2008

  8. [8]

    Toward storytelling from visual lifelogging: An overview

    Bolanos, M., Dimiccoli, M., and Radeva, P. Toward storytelling from visual lifelogging: An overview. IEEE Transactions on Human-Machine Systems, 47 0 (1): 0 77--90, 2016

    work page 2016

Show all 77 references
  1. [9]

    Yolo-hand-detection

    Cansik. Yolo-hand-detection. https://github.com/cansik/yolo-hand-detection, 2020

    2020

  2. [10]

    When i look into your eyes: A survey on computer vision contributions for human gaze estimation and tracking

    Cazzato, D., Leo, M., Distante, C., and Voos, H. When i look into your eyes: A survey on computer vision contributions for human gaze estimation and tracking. Sensors, 20 0 (13): 0 3739, 2020

    2020

  3. [11]

    Chakraborty, A., Mandal, B., and Galoogahi, H. K. Person re-identification using multiple first-person-views on wearable devices. In WACV, 2016

    2016

  4. [12]

    where am i?

    Chen, J., Barath, D., Armeni, I., Pollefeys, M., and Blum, H. “where am i?” scene retrieval with language. In European Conference on Computer Vision, pp.\ 201--220. Springer, 2024

    2024

  5. [13]

    Learning a dynamic privacy-preserving camera robust to inversion attacks

    Cheng, J., Dai, X., Wan, J., Antipa, N., and Vasconcelos, N. Learning a dynamic privacy-preserving camera robust to inversion attacks. In ECCV, 2024 a

    2024

  6. [14]

    Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms

    Cheng, Z., Leng, S., Zhang, H., Xin, Y., Li, X., Chen, G., Zhu, Y., Zhang, W., Luo, Z., Zhao, D., et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024 b

    2024 arXiv

  7. [15]

    Object removal by exemplar-based inpainting

    Criminisi, A., Perez, P., and Toyama, K. Object removal by exemplar-based inpainting. In CVPR, 2003

    2003

  8. [16]

    Region filling and object removal by exemplar-based image inpainting

    Criminisi, A., P \'e rez, P., and Toyama, K. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing (TIP), 13 0 (9): 0 1200--1212, 2004

    2004

  9. [17]

    G., Tan, C., Lim, J.-H., and Tan, A.-H

    Del Molino, A. G., Tan, C., Lim, J.-H., and Tan, A.-H. Summarization of egocentric videos: A comprehensive survey. IEEE Transactions on Human-Machine Systems, 47 0 (1): 0 65--76, 2016

    2016

  10. [18]

    Retinaface: Single-stage dense face localisation in the wild, 2019

    Deng, J., Guo, J., Zhou, Y., Yu, J., Kotsia, I., and Zafeiriou, S. Retinaface: Single-stage dense face localisation in the wild, 2019. URL https://arxiv.org/abs/1905.00641

    2019 arXiv

  11. [19]

    Mitigating bystander privacy concerns in egocentric activity recognition with deep learning and intentional image degradation

    Dimiccoli, M., Mar \' n, J., and Thomaz, E. Mitigating bystander privacy concerns in egocentric activity recognition with deep learning and intentional image degradation. ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (IMWUT), 2018

    2018

  12. [20]

    From third person to first person: Dataset and baselines for synthesis and retrieval

    Elfeki, M., Regmi, K., Ardeshir, S., and Borji, A. From third person to first person: Dataset and baselines for synthesis and retrieval. arXiv preprint arXiv:1812.00104, 2018

    2018 arXiv

  13. [21]

    M., Jain, A., Schmidt, L., Toshev, A., and Shankar, V

    Fang, A., Jose, A. M., Jain, A., Schmidt, L., Toshev, A., and Shankar, V. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023

    2023 arXiv

  14. [22]

    and Oni, V

    Farringdon, J. and Oni, V. Visual augmented memory (vam). In Digest of Papers. Fourth International Symposium on Wearable Computers, pp.\ 167--168. IEEE, 2000

    2000

  15. [23]

    K., and Rehg, J

    Fathi, A., Hodgins, J. K., and Rehg, J. M. Social interactions: A first-person perspective. In CVPR, 2012

    2012

  16. [24]

    Body part based re-identification from an egocentric perspective

    Fergnani, F., Alletto, S., Serra, G., De Mira, J., and Cucchiara, R. Body part based re-identification from an egocentric perspective. In CVPR Workshops, 2016

    2016

  17. [25]

    K., Ryan, F., Sharma, J., Wray, M., Xu, M., Xu, E

    Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., Martin, M., Nagarajan, T., Radosavovic, I., Ramakrishnan, S. K., Ryan, F., Sharma, J., Wray, M., Xu, M., Xu, E. Z., Zhao, C., Bansal, S., Batra, D., Cartilli...

    2022

  18. [26]

    M., Jain, S., Khirodkar, R., Kukreja, D., Liang, K

    Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., Byrne, E., Chavis, Z., Chen, J., Cheng, F., Chu, F.-J., Crane, S., Dasgupta, A., Dong, J., Escobar, M., Forigua, C., Gebreselasie, A., Haresh, S., Hua...

    2024

  19. [27]

    Gurari, D., Li, Q., Lin, C., Zhao, Y., Guo, A., Stangl, A., and Bigham, J. P. Vizwiz-priv: A dataset for recognizing the presence and purpose of private visual information in images taken by blind people. In CVPR, 2019

    2019

  20. [28]

    T., et al

    Hasan, R., Shaffer, P., Crandall, D., Apu Kapadia, E. T., et al. Cartooning for enhanced privacy in lifelogging and streaming videos. In CVPR Workshops, pp.\ 29--38, 2017

    2017

  21. [29]

    C., and Arguello, H

    Hinojosa, C., Niebles, J. C., and Arguello, H. Learning privacy-preserving optics for human pose estimation. In ICCV, 2021

    2021

  22. [30]

    Hinojosa, C., Marquez, M., Arguello, H., Adeli, E., Fei-Fei, L., and Niebles, J. C. Privhar: Recognizing human actions from privacy-preserving lens. In ECCV, 2022

    2022

  23. [31]

    and Peleg, S

    Hoshen, Y. and Peleg, S. An egocentric look at video photographer identity. In CVPR, 2016

    2016

  24. [32]

    Privacy behaviors of lifeloggers using wearable cameras

    Hoyle, R., Templeman, R., Armes, S., Anthony, D., Crandall, D., and Kapadia, A. Privacy behaviors of lifeloggers using wearable cameras. In Proceedings of the 2014 ACM international joint conference on pervasive and ubiquitous computing, pp.\ 571--582, 2014

    2014

  25. [33]

    Sensitive Lifelogs: A Privacy Analysis of Photos from Wearable Cameras

    Hoyle, R., Templeman, R., Anthony, D., Crandall, D., and Kapadia, A. Sensitive Lifelogs: A Privacy Analysis of Photos from Wearable Cameras . In Conference on Human Factors in Computing Systems, 2015 a

    2015

  26. [34]

    Sensitive lifelogs: A privacy analysis of photos from wearable cameras

    Hoyle, R., Templeman, R., Anthony, D., Crandall, D., and Kapadia, A. Sensitive lifelogs: A privacy analysis of photos from wearable cameras. In Proceedings of the 33rd Annual ACM conference on human factors in computing systems, pp.\ 1645--1648, 2015 b

    2015

  27. [35]

    and Joo, J

    Karkkainen, K. and Joo, J. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp.\ 1548--1558, 2021

    2021

  28. [36]

    S., Yu, X., Mitra, K., Chandraker, M., and Pittaluga, F

    Khan, S. S., Yu, X., Mitra, K., Chandraker, M., and Pittaluga, F. Opencam: Lensless optical encryption camera. IEEE Transactions on Computational Imaging, 2024

    2024

  29. [37]

    Supervised contrastive learning

    Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., and Krishnan, D. Supervised contrastive learning. NeurIPS, 2020

    2020

  30. [38]

    Enhancing lifelogging privacy by detecting screens

    Korayem, M., Templeman, R., Chen, D., Crandall, D., and Kapadia, A. Enhancing lifelogging privacy by detecting screens. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, pp.\ 4309--4314, 2016

    2016

  31. [39]

    A wearable face recognition system for individuals with visual impairments

    Krishna, S., Little, G., Black, J., and Panchanathan, S. A wearable face recognition system for individuals with visual impairments. In Proceedings of the 7th international ACM SIGACCESS conference on Computers and accessibility, pp.\ 106--113, 2005

    2005

  32. [40]

    Ego-exo: Transferring visual representations from third-person to first-person videos

    Li, Y., Nagarajan, T., Xiong, B., and Grauman, K. Ego-exo: Transferring visual representations from third-person to first-person videos. In CVPR, 2021

    2021

  33. [41]

    Exocentric to egocentric image generation via parallel generative adversarial network

    Liu, G., Tang, H., Latapie, H., and Yan, Y. Exocentric to egocentric image generation via parallel generative adversarial network. In ICASSP, 2020

    2020

  34. [42]

    M., Corso, J

    Liu, G., Tang, H., Latapie, H. M., Corso, J. J., and Yan, Y. Cross-view exocentric to egocentric video synthesis. In ACM International Conference on Multimedia, 2021

    2021

  35. [43]

    Liu, H., Li, C., Wu, Q., and Lee, Y. J. Visual instruction tuning. Advances in neural information processing systems, 36, 2024 a

    2024

  36. [44]

    Towards balanced alignment: Modal-enhanced semantic modeling for video moment retrieval

    Liu, Z., Li, J., Xie, H., Li, P., Ge, J., Liu, S.-A., and Jin, G. Towards balanced alignment: Modal-enhanced semantic modeling for video moment retrieval. In AAAI, 2024 b

    2024

  37. [45]

    Zero-shot video moment retrieval from frozen vision-language models

    Luo, D., Huang, J., Gong, S., Jin, H., and Liu, Y. Zero-shot video moment retrieval from frozen vision-language models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp.\ 5464--5473, 2024 a

    2024

  38. [46]

    Intention-driven ego-to-exo video generation

    Luo, H., Zhu, K., Zhai, W., and Cao, Y. Intention-driven ego-to-exo video generation. arXiv preprint arXiv:2403.09194, 2024 b

    2024 arXiv

  39. [47]

    Put myself in your shoes: Lifting the egocentric perspective from exocentric videos

    Luo, M., Xue, Z., Dimakis, A., and Grauman, K. Put myself in your shoes: Lifting the egocentric perspective from exocentric videos. arXiv preprint arXiv:2403.06351, 2024 c

    2024 arXiv

  40. [48]

    A wearable face recognition system on google glass for assisting social interactions

    Mandal, B., Chia, S.-C., Li, L., Chandrasekhar, V., Tan, C., and Lim, J.-H. A wearable face recognition system on google glass for assisting social interactions. In ACCV Workshops, 2014

    2014

  41. [49]

    Matkowski, W. M. and Kong, A. W. K. Gender and ethnicity classification based on palmprint and palmar hand images from uncontrolled environment. In IJCB. IEEE, 2020

    2020

  42. [50]

    M., Chai, T., and Kong, A

    Matkowski, W. M., Chai, T., and Kong, A. W. K. Palmprint recognition in uncontrolled and uncooperative environment. IEEE Transactions on Information Forensics and Security (TIFS), 15: 0 1601--1615, 2019

    2019

  43. [51]

    Audio-visual instance discrimination with cross-modal agreement

    Morgado, P., Vasconcelos, N., and Misra, I. Audio-visual instance discrimination with cross-modal agreement. In CVPR, 2021

    2021

  44. [52]

    Recognition of activities of daily living with egocentric vision: A review

    Nguyen, T.-H.-C., Nebel, J.-C., and Florez-Revuelta, F. Recognition of activities of daily living with egocentric vision: A review. Sensors, 16 0 (1): 0 72, 2016

    2016

  45. [53]

    Oord, A. v. d., Li, Y., and Vinyals, O. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  46. [54]

    Towards a visual privacy advisor: Understanding and predicting privacy risks in images

    Orekondy, T., Schiele, B., and Fritz, M. Towards a visual privacy advisor: Understanding and predicting privacy risks in images. In ICCV, 2017

    2017

  47. [55]

    M., Damen, D., and Tommasi, T

    Plizzari, C., Goletto, G., Furnari, A., Bansal, S., Ragusa, F., Farinella, G. M., Damen, D., and Tommasi, T. An outlook into the future of egocentric vision. International Journal of Computer Vision, pp.\ 1--57, 2024

    2024

  48. [56]

    Head motion signatures from egocentric videos

    Poleg, Y., Arora, C., and Peleg, S. Head motion signatures from egocentric videos. In ACCV, 2015

    2015

  49. [57]

    Q., Shah, H., Shou, M

    Pramanick, S., Song, Y., Nag, S., Lin, K. Q., Shah, H., Shou, M. Z., Chellappa, R., and Zhang, P. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In ICCV, 2023

    2023

  50. [58]

    A., Stuart, A., Calikli, G., Mccormick, C., Mehta, V., Hutton, L., Bandara, A

    Price, B. A., Stuart, A., Calikli, G., Mccormick, C., Mehta, V., Hutton, L., Bandara, A. K., Levine, M., and Nuseibeh, B. Logging you, Logging me: A Replicable Study of Privacy and Sharing Behaviour in Groups of Visual Lifeloggers . ACM on Interactive, Mobile, Wearable and Ubi...

    2017

  51. [59]

    P.-W., Gu, X., Jobarteh, M

    Qiu, J., Lo, F. P.-W., Gu, X., Jobarteh, M. L., Jia, W., Baranowski, T., Steiner-Asiedu, M., Anderson, A. K., McCrory, M. A., Sazonov, E., et al. Egocentric image captioning for privacy-preserved passive dietary intake monitoring. IEEE Transactions on Cybernetics, 54 0 (2): 0 ...

    2023

  52. [60]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In ICML, 2021

    2021

  53. [61]

    J., and Ryoo, M

    Ren, Z., Lee, Y. J., and Ryoo, M. S. Learning to anonymize faces for privacy preserving action detection. In ECCV, 2018

    2018

  54. [62]

    Ryoo, M., Rothrock, B., Fleming, C., and Yang, H. J. Privacy-preserving human activity recognition from extreme low resolution. In AAAI, 2017

    2017

  55. [63]

    A., Gupta, A., Schmid, C., Farhadi, A., and Alahari, K

    Sigurdsson, G. A., Gupta, A., Schmid, C., Farhadi, A., and Alahari, K. Actor and observer: Joint modeling of first and third-person videos. In CVPR, 2018 a

    2018

  56. [64]

    A., Gupta, A., Schmid, C., Farhadi, A., and Alahari, K

    Sigurdsson, G. A., Gupta, A., Schmid, C., Farhadi, A., and Alahari, K. Charades-ego: A large-scale dataset of paired third and first person videos. arXiv preprint arXiv:1804.09626, 2018 b

    2018 arXiv

  57. [65]

    L., Kang, S

    Speciale, P., Schönberger, J. L., Kang, S. B., Sinha, S. N., and Pollefeys, M. Privacy Preserving Image-Based Localization . In CVPR, 2019

    2019

  58. [66]

    J., and Kapadia, A

    Templeman, R., Korayem, M., Crandall, D. J., and Kapadia, A. Placeavoider: Steering first-person cameras away from sensitive spaces. In NDSS, 2014

    2014

  59. [67]

    Is sharing of egocentric video giving away your biometric signature? In ECCV, 2020 a

    Thapar, D., Arora, C., and Nigam, A. Is sharing of egocentric video giving away your biometric signature? In ECCV, 2020 a

    2020

  60. [68]

    Recognizing camera wearer from hand gestures in egocentric videos

    Thapar, D., Nigam, A., and Arora, C. Recognizing camera wearer from hand gestures in egocentric videos. In International Conference on Multimedia, 2020 b

    2020

  61. [69]

    Anonymizing egocentric videos

    Thapar, D., Nigam, A., and Arora, C. Anonymizing egocentric videos. In ICCV, 2021

    2021

  62. [70]

    and Kovashka, A

    Thomas, C. and Kovashka, A. Seeing behind the camera: Identifying the authorship of a photograph. In CVPR, 2016

    2016

  63. [71]

    Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training

    Tong, Z., Song, Y., Wang, J., and Wang, L. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35: 0 10078--10093, 2022

    2022

  64. [72]

    Tsutsui, S., Fu, Y., and Crandall, D. J. Whose hand is this? person identification from egocentric hand gestures. In WACV, 2021

    2021

  65. [73]

    Retrieval-augmented egocentric video captioning

    Xu, J., Huang, Y., Hou, J., Chen, G., Zhang, Y., Feng, R., and Xie, W. Retrieval-augmented egocentric video captioning. In CVPR, 2024

    2024

  66. [74]

    M., and Sato, Y

    Yonetani, R., Kitani, K. M., and Sato, Y. Ego-surfing first-person videos. In CVPR, pp.\ 5445--5454, 2015

    2015

  67. [75]

    First-and third-person video co-analysis by learning spatial-temporal joint attention

    Yu, H., Cai, M., Liu, Y., and Lu, F. First-and third-person video co-analysis by learning spatial-temporal joint attention. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45 0 (6): 0 6631--6646, 2020

    2020

  68. [76]

    Beyond frontal faces: Improving person recognition using multiple cues

    Zhang, N., Paluri, M., Taigman, Y., Fergus, R., and Bourdev, L. Beyond frontal faces: Improving person recognition using multiple cues. In CVPR, 2015

    2015

  69. [77]

    Fusing personal and environmental cues for identification and segmentation of first-person camera wearers in third-person views

    Zhao, Z., Wang, Y., and Wang, C. Fusing personal and environmental cues for identification and segmentation of first-person camera wearers in third-person views. In CVPR, 2024

    2024

Pith tools

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