Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Speaker Embedding Informed Audiovisual Active Speaker Detection for Egocentric Recordings

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that its SCAN module, which compares reference-speech and candidate-audio speaker embeddings through cross-attention, improves egocentric active speaker detection by 14.5% and 10.3% relative mAP over TalkNet and…

desk verdict SCAN is a plausible auxiliary module with consistent but unverified gains; the protocol must rule out identical-waveform matching before the numbers can be trusted. read the letter →

arxiv 2502.06012 v1 pith:PFEVFR6E submitted 2025-02-09 cs.MM

Jason Clarke , Yoshihiko Gotoh , Stefan Goetze This is my paper · ORCID
classification cs.MM
keywords activespeakerdetectionembeddingegocentricvideocross-attentionself-supervisedfacerecognitionidentity-speechlibraryEgo4Daudiovisualfusion
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that active speaker detection on egocentric video becomes substantially more reliable when the system can compare the candidate speaker's reference voice against the current audio moment by moment. The proposed SCAN module extracts speaker embeddings from both pre-enrolled reference speech and the candidate audio track, then uses cross-attention to highlight when the current audio matches the reference voice, so frames with occluded faces or competing speakers can still be attributed correctly. Plugged into TalkNet and Light-ASD, SCAN raises mean average precision on the Ego4D-AVD validation fold by 14.5% and 10.3%, respectively, nearly matching larger recent models. The paper also contributes a self-supervised method that finetunes a face recognition model on video tracks to build identity-speech libraries robust to egocentric noise, replacing the prior enrollment approach. If these results hold, lightweight auxiliary speaker comparison could become a standard add-on for ASD systems deployed on wearable devices.

What carries the argument

SCAN is a lightweight auxiliary network that, for each video frame, takes a one-second window of the candidate audio track and a clip of reference speech from the identity-speech library, embeds both with a frozen ECAPA-TDNN speaker verifier, and applies cross-attention so the audio window queries the reference clip. The attended output is classified as active or inactive and added as an auxiliary binary cross-entropy loss to the baseline ASD system. The second mechanism is a self-supervised face recognition finetune: an existing frame-based face model is wrapped with transformer encoder layers and trained to spot randomly inserted impostor frames within a face track, which teaches it to weight crisp identity-bearing frames and yields a track-level identity embedding for enrolling the speech library.

What would settle it

Take the Ego4D validation clips where the candidate's face is occluded and another voice is active; if TalkNet+SCAN does not beat TalkNet on that subset, the specific disambiguation claim fails.

Watch

Extended reading notes

Core claim

The authors claim that the limiting factor for speaker-embedding-informed ASD is not just having reference speech, but actively comparing it with the candidate audio signal at frame level. SCAN does this by computing a cross-attention between embeddings of one-second audio windows from the test track and embeddings of reference speech from the identity-speech library, with the frozen ECAPA-TDNN speaker encoder. The resulting comparison signal is fed through a binary classifier as an auxiliary loss, and the output is shown to transfer to two very different baselines: TalkNet+SCAN reaches 58.4 mAP on Ego4D with a hypothesized library (58.0 without), and Light-ASD+SCAN reaches 59.9 (57.1 without); against the previous TS-TalkNet approach, TalkNet+SCAN is 4.4 points better with a hypothesized library. On AVA-ActiveSpeaker the gains are small, which the authors attribute to the rarity of visually occluded multi-talker scenes in exocentric footage. A second claimed contribution is that identity-speech libraries built with the self-supervised temporal face model are better than TS-TalkNet's enrollment, moving the cluster-separability score from -0.17 to 0.16 on Ego4D validation.

Load-bearing premise

The claimed gains depend on the stored reference clips for each face actually being that person's voice, and the paper's own evaluation reports only minor separability between same-person and different-person clips (a score of 0.16), so enrollment noise is the point of fragility.

Editorial extensions

If this is right

  • ASD systems that ignore speaker identity can be upgraded with a small auxiliary module rather than redesigned, since SCAN is architecture-agnostic.
  • On egocentric benchmarks, speaker-embedding-informed ASD can approach the performance of much larger specialized models: Light-ASD+SCAN reaches 59.9 mAP on Ego4D versus 60.7 for the current listed top system.
  • The benefit concentrates in exactly the hard cases of visual occlusion and overlapping speech, so gains on exocentric data should stay small, as observed on AVA-ActiveSpeaker.
  • Identity-speech library quality is a bottleneck: better enrollment directly raises the ceiling, as shown by the gap between ground-truth and hypothesized libraries for TalkNet+SCAN (58.4 vs 58.0) and Light-ASD+SCAN (59.9 vs 57.1).

Reading between the lines

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

  • The cross-attention comparison could be applied to any ASD backbone with an audio encoder, and its benefit should scale with the degree of visual corruption; a natural next test is a dataset built specifically from wearable cameras with head motion.
  • Because SCAN's reference audio is drawn from a pre-enrolled library, the method is compatible with a streaming setting where reference speech is accumulated after the first confirmed utterance, turning ASD into a bootstrapping process.
  • The reported cluster-separability score of 0.16 suggests the enrollment model still confuses some identities; improving temporal face clustering, or adding a speaker-verification filter on enrolled speech, would likely push the Ego4D mAP above the reported numbers.
  • One implication the authors do not draw out: SCAN's framewise comparison output could be reused as a weak diarization signal, since high attention between candidate audio and reference speech marks who is speaking at each moment.
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

3 major / 5 minor

Summary. The paper proposes SCAN, an auxiliary module for active speaker detection (ASD) that extracts speaker embeddings from both reference speech and the candidate audio signal, and compares them via a cross-attention mechanism. The authors also propose a self-supervised method for building identity-speech libraries by fine-tuning a face recognition model on video tracks with transformer layers and an impostor-frame classification objective. Experiments on Ego4D-AVD show that adding SCAN improves two speaker-embedding-naive baselines, TalkNet and Light-ASD, by 14.5% and 10.3% relative mAP when ground-truth identity-speech libraries are used, with smaller gains on AVA-ActiveSpeaker. The identity-speech library quality is evaluated using silhouette scores.

Significance. If the reported results hold, SCAN is a lightweight and modular contribution that can be attached to existing ASD systems and improves robustness in egocentric recordings. The use of a frozen ECAPA-TDNN encoder and evaluation on two public benchmarks are strengths. The idea of frame-level cross-attention between reference speech and candidate audio is interesting and distinct from prior speaker-embedding injection. However, the central claim is currently weakened by two issues: the paper does not rule out temporal overlap between reference speech and the audio being scored, and the evaluation lacks error bars or statistical tests. The identity-speech library evaluation is also thin, resting on a single silhouette score that the authors acknowledge shows only minor cluster separability. The contribution is therefore promising but not yet convincingly established.

major comments (3)
  1. [Section III-B and Eq. (1)] The paper does not state whether reference speech A_S is temporally disjoint from the candidate audio A during training and evaluation. Since the identity-speech library is populated with pre-diarised speech from the tracks themselves (Section II-C1), A_S can be drawn from the same recording, or even the same waveform, as the audio window being scored. In that case, f_phi(A_S) and f_phi(A) are embeddings of identical acoustic content, and the cross-attention in Eq. (1) can peak at self-matching frames, allowing the classifier to rely on acoustic self-similarity rather than speaker identity. This is a load-bearing issue because it could inflate the Ego4D mAP gains in Table I. Please specify the temporal/excerpt exclusion rule used when sampling A_S, and add an experiment in which enrollment speech is guaranteed to come from a different recording or a non-overlapping time interval.
  2. [Section IV-A and Table I] The claim that SCAN "significantly improves" the baselines is not supported by error bars, multiple runs, or significance tests. The AVA differences are small (for example, 93.9 vs 94.2 for Light-ASD with a ground-truth library), and without variability estimates it is not possible to tell which reported gaps are meaningful. Please report standard deviations over at least three seeds or otherwise provide a statistical comparison.
  3. [Section IV-B and Table III] The evaluation of the proposed identity-speech library method rests on a single silhouette score (0.16), which the authors themselves describe as "only minor cluster separability". This does not support the conclusion of a "substantial improvement" over TS-TalkNet's library generation. In particular, there is no downstream comparison of ASD mAP using the same baseline with the previous library versus the proposed library. Please add such a comparison, and consider reporting precision/recall of identity clustering against the pseudo-identity annotations mentioned in Section III-A.
minor comments (5)
  1. [Abstract and Table II] The abstract and Table II report "14.5% and 10.3%" improvements without specifying that they correspond to ground-truth identity-speech libraries; with hypothesised libraries the relative gains are 13.7% and 5.2%. Please state the library condition explicitly.
  2. [Section II-C1] The static cosine similarity threshold (0.9) and the minimum reference-speech duration (2.5 s) appear to be ad hoc choices; a sensitivity analysis would make the identity-speech library method more reproducible and help assess its robustness.
  3. [Section III-C] Validation-only evaluation is a standard limitation in ASD, but please state this in the abstract or conclusion so readers do not mistake the reported numbers for test-fold results.
  4. [Figure 4] Figure 4 lacks axis labels and a legend; adding them would make the similarity distributions interpretable.
  5. [Table I and Table II] Some table rows are hard to parse; for example, the row grouping in Table I is not fully clear from the header, and Table II cites SPELL via reference [28], which is a technical report on long-term action anticipation rather than an ASD system. Please check the citation and clarify the row semantics.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity: SCAN's Ego4D gains are direct ablations on identical baselines; only a minor non-load-bearing self-citation [9] is present.

  1. other [I. Introduction (ref. [9]); III-C Evaluation Metric (ref. [9])]
    "This problem is also demonstrated by the disparity in performance when evaluating ASD systems on exocentric [8] vs egocentric [9], [14], [15] benchmarks, where in the latter, challenging scenes are regularly prevalent."

    Reference [9] is the authors' own ASRU23 paper. It appears only to motivate egocentric-recording challenges and to cite the field convention of reporting on validation folds; neither usage supplies the SCAN architecture, the identity-library method, or the baseline numbers in the central claim. The central mAP gains (14.5% TalkNet, 10.3% Light-ASD) come from in-paper ablations on identical baselines with a frozen ECAPA-TDNN encoder, so the self-citation is not load-bearing and the claim is not forced by it. Flagged for completeness per the review rule; it does not constitute circularity.

full rationale

The derivation chain is self-contained. SCAN's output (Eq. 1) is a softmax-weighted combination of frozen ECAPA-TDNN embeddings of the candidate audio (queries) and reference speech (keys/values); the framewise classifier is trained with binary cross-entropy against the dataset's ground-truth activity labels. The label is never an input to Eq. (1), so the prediction does not reduce to its inputs by construction (no self-definition). Identity-speech libraries are built independently via face-embedding clustering (threshold 0.9) plus pre-diarised speech and are evaluated separately with a silhouette score on the validation fold; the paper candidly reports 0.16 as 'only minor cluster separability', a limitation that weakens deployment claims but does not make the ASD comparison circular. Baselines TalkNet/Light-ASD are reimplemented with their original settings and compared with and without SCAN under an identical protocol; the 14.5%/10.3% relative mAP gains are direct ablations, not fitted parameters renamed as predictions. External comparisons (LoCoNet, SPELL) come from original manuscripts, not self-citation. The only self-referential element is reference [9] (authors' ASRU23 paper), cited for egocentric-challenge motivation and the validation-fold evaluation convention; it is not load-bearing. One caveat, noted as a correctness risk rather than circularity: Section III-B samples reference speech 'randomly selected from the relevant part of the identity-speech library' and Section II-C1 builds that library from 'all tracks in a given dataset', without stating temporal exclusion from the scored audio A. If AS is drawn from the same clip as A, Eq. (1) could match identical waveforms; this is a potential evaluation-leakage risk the paper does not rule out, but the paper does not define AS in terms of A, so no definitional reduction exists and the concern belongs in a correctness review, not in the circularity score.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-chosen hyperparameters for library construction and SCAN, on the domain assumptions that tracks are identity-homogeneous and that frozen speaker embeddings remain discriminative in ego noise, and on the assumption that the self-supervised face finetuning generalizes. No new physical or ontological entities are postulated.

free parameters (6)
  • identity cosine similarity threshold = 0.9
    Used in Section II-C1 to decide whether two face tracks belong to the same identity; hand-chosen and not optimized; directly affects library purity.
  • minimum reference speech duration = 2.5 s
    Section III-B; speech shorter than this is excluded from the identity-speech library; hand-chosen.
  • SCAN audio window length = 1 s
    Section III-B; each framewise speaker embedding is extracted from a 1-second window of candidate audio; hand-chosen.
  • SCAN output embedding dimension = 64
    Section III-B; dimensionality of FS after cross-attention; architecture choice.
  • impostor frame insertion rate = 30%
    Section III-B; fraction of impostor frames polluting tracks during face-recognition finetuning; hand-chosen.
  • number of transformer encoder layers = 4
    Section III-B; L=4 layers in the self-supervised face model; architectural choice.
assumptions (4)
  • domain assumption ASD dataset tracks are identity-homogeneous, so clustering per-track face embeddings yields identity clusters and corresponding speech.
    Section II-C: 'by definition, tracks are identity-homogeneous'; this is the foundation of identity-speech library construction.
  • domain assumption Frozen ECAPA-TDNN speaker embeddings remain informative when computed from 1-second noisy egocentric audio windows.
    Section II-B and III-B; if embeddings are unreliable under ego noise, SCAN's cross-attention input is uninformative.
  • domain assumption The self-supervised native/impostor classification objective teaches the face model to focus on clean native frames and ignore distorted or occluded frames.
    Section II-C2; the paper argues the model assigns low weighting to non-relevant frames; this generalization is assumed, not proven.
  • domain assumption Validation folds of Ego4D-AVD and AVA-ActiveSpeaker are representative for comparing ASD systems.
    Section III-C; test fold labels are unavailable, so all comparisons are on validation folds; standard but an assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Speaker Embedding Informed Audiovisual Active Speaker Detection for Egocentric Recordings." pith.science (2026). https://pith.science/paper/PFEVFR6E

@misc{pith2026250206012,
  author       = {Pith},
  title        = {Pith review of: Speaker Embedding Informed Audiovisual Active Speaker Detection for Egocentric Recordings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PFEVFR6E}},
  note         = {Machine review of arXiv:2502.06012}
}
read the original abstract

Audiovisual active speaker detection (ASD) addresses the task of determining the speech activity of a candidate speaker given acoustic and visual data. Typically, systems model the temporal correspondence of audiovisual cues, such as the synchronisation between speech and lip movement. Recent work has explored extending this paradigm by additionally leveraging speaker embeddings extracted from candidate speaker reference speech. This paper proposes the speaker comparison auxiliary network (SCAN) which uses speaker-specific information from both reference speech and the candidate audio signal to disambiguate challenging scenes when the visual signal is unresolvable. Furthermore, an improved method for enrolling face-speaker libraries is developed, which implements a self-supervised approach to video-based face recognition. Fitting with the recent proliferation of wearable devices, this work focuses on improving speaker-embedding-informed ASD in the context of egocentric recordings, which can be characterised by acoustic noise and highly dynamic scenes. SCAN is implemented with two well-established baselines, namely TalkNet and Light-ASD; yielding a relative improvement in mAP of 14.5% and 10.3% on the Ego4D benchmark, respectively.

Figures

Figures reproduced from arXiv: 2502.06012 by the authors.

Figure 1
Figure 1. Example of typical false-positive ASD: a) input audio signal; b) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. SCAN is shown in the top box which leverages speaker-specific [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Self-supervised video-based face recognition model. impostor [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Similarity between same-identity embeddings and different-identity [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 13 canonical work pages

  1. [1]

    Is Someone Speaking? Exploring Long-term Temporal Features for Audio-visual Active Speaker Detection,

    R. Tao, Z. Pan, R. K. Das, X. Qian, M. Z. Shou, and H. Li, “Is Someone Speaking? Exploring Long-term Temporal Features for Audio-visual Active Speaker Detection,” in Proc. 29th ACM Int. Conf. on Multimedia , 2021

    work page 2021

  2. [2]

    Combining Residual Networks with LSTMs for Lipreading,

    T. Stafylakis and G. Tzimiropoulos, “Combining Residual Networks with LSTMs for Lipreading,” in Proc. Interspeech 2017 , 2017

    work page 2017

  3. [3]

    Active Speakers in Context,

    J. L. Alcazar, F. C. Heilbron, L. Mai, F. Perazzi, J.-Y . Lee, P. Arbel ´aez, and B. Ghanem, “Active Speakers in Context,” 2020 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) , 2020

    work page 2020

  4. [4]

    ASD-Transformer: Efficient Active Speaker Detection Using Self And Multimodal Transformers,

    G. Datta, T. Etchart, V . Yadav, V . Hedau, P. Natarajan, and S.-F. Chang, “ASD-Transformer: Efficient Active Speaker Detection Using Self And Multimodal Transformers,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP) , 2022

    2022

  5. [5]

    Hello! My name is... Buffy

    M. Everingham, J. Sivic, and A. Zisserman, “Hello! My name is... Buffy” – Automatic Naming of Characters in TV Video,” in British Machine Vision Conference, 2006

    2006

  6. [6]

    MAAS: Multi-modal Assignation for Active Speaker Detection,

    J. Le’on-Alc’azar, F. C. Heilbron, A. K. Thabet, and B. Ghanem, “MAAS: Multi-modal Assignation for Active Speaker Detection,” 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , 2021

    2021

  7. [7]

    Target Active Speaker Detection with Audio-visual Cues,

    Y . Jiang, R. Tao, Z. Pan, and H. Li, “Target Active Speaker Detection with Audio-visual Cues,” in Proc. Interspeech, 2023

    2023

  8. [8]

    Ava Active Speaker: An Audio-Visual Dataset for Active Speaker De- tection,

    J. Roth, S. Chaudhuri, O. Klejch, R. Marvin, A. Gallagher, L. Kaver, S. Ramaswamy, A. Stopczynski, C. Schmid, Z. Xi, and C. Pantofaru, “Ava Active Speaker: An Audio-Visual Dataset for Active Speaker De- tection,” in Proc. Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP), 2020

    2020

Show all 28 references
  1. [9]

    Improving Audiovisual Active Speaker Detection in Egocentric Recordings with the Data-Efficient Image Transformer,

    J. Clarke, Y . Gotoh, and S. Goetze, “Improving Audiovisual Active Speaker Detection in Egocentric Recordings with the Data-Efficient Image Transformer,” in IEEE Automatic Speech Recognition and Un- derstanding Workshop (ASRU23) , 2023

    2023

  2. [10]

    Ego4D: Around the World in 3,000 Hours of Egocentric Video,

    “Ego4D: Around the World in 3,000 Hours of Egocentric Video,” 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2022

  3. [11]

    End-to-End Active Speaker Detection, author=Juan Leon Alcazar and Moritz Cordes and Chen Zhao and Bernard Ghanem,

    “End-to-End Active Speaker Detection, author=Juan Leon Alcazar and Moritz Cordes and Chen Zhao and Bernard Ghanem,” in European Conference on Computer Vision , 2022

    2022

  4. [12]

    How to Design a Three-Stage Architecture for Audio-Visual Active Speaker Detection in the Wild,

    O. K ¨op¨ukl¨u, M. Taseska, and G. Rigoll, “How to Design a Three-Stage Architecture for Audio-Visual Active Speaker Detection in the Wild,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , 2021

    2021

  5. [13]

    A Light Weight Model for Active Speaker Detection,

    J. Liao, H. Duan, K. Feng, W. Zhao, Y . Yang, and L. Chen, “A Light Weight Model for Active Speaker Detection,” in Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , June 2023

    2023

  6. [14]

    LoCoNet: Long-Short Context Network for Active Speaker Detection,

    X. Wang, F. Cheng, and G. Bertasius, “LoCoNet: Long-Short Context Network for Active Speaker Detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , June 2024

    2024

  7. [15]

    Learning Long-Term Spatial-Temporal Graphs for Active Speaker Detection,

    K. Min, S. Roy, S. Tripathi, T. Guha, and S. Majumdar, “Learning Long-Term Spatial-Temporal Graphs for Active Speaker Detection,” in Euro. Conf. on Computer Vision , 2022

    2022

  8. [16]

    ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification,” in Interspeech 2020 . ISCA, Oct 2020. [Online]. Available: http://dx.doi.org/10.21437/Interspeech.2020-2650

    2020 doi

  9. [17]

    [Online]

    “Personal V AD 2.0: Optimizing Personal V oice Activity Detection for On-Device Speech Recognition, author=Shaojin Ding and Rajeev Rikhye and Qiao Liang and Yanzhang He and Quan Wang and Arun Narayanan and Tom O’Malley and Ian McGraw,” 2022. [Online]. Available: https://arxiv....

    2022 arXiv

  10. [18]

    Efficient Personal V oice Activity Detection with Wake Word Reference Speech,

    B. Zeng, M. Cheng, Y . Tian, H. Liu, and M. Li, “Efficient Personal V oice Activity Detection with Wake Word Reference Speech,” inICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 12 241–12 245

    2024

  11. [19]

    V oxCeleb: A Large-Scale Speaker Identification Dataset,

    A. Nagrani, J. S. Chung, and A. Zisserman, “V oxCeleb: A Large-Scale Speaker Identification Dataset,” in Interspeech 2017. ISCA, Aug. 2017. [Online]. Available: http://dx.doi.org/10.21437/Interspeech.2017-950

    2017 doi

  12. [20]

    Ring Loss: Convex Feature Normalization for Face Recognition,

    Y . Zheng, D. K. Pal, and M. Savvides, “Ring Loss: Convex Feature Normalization for Face Recognition,” 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2018

    2018

  13. [21]

    ArcFace: Additive Angular Margin Loss for Deep Face Recognition,

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “ArcFace: Additive Angular Margin Loss for Deep Face Recognition,” in2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2019

    2019

  14. [22]

    SphereFace: Deep Hypersphere Embedding for Face Recognition,

    W. Liu, Y . Wen, Z. Yu, M. Li, B. Raj, and L. Song, “SphereFace: Deep Hypersphere Embedding for Face Recognition,” 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017

    2017

  15. [23]

    CosFace: Large Margin Cosine Loss for Deep Face Recognition,

    H. Wang, Y . Wang, Z. Zhou, X. Ji, Z. Li, D. Gong, J. Zhou, and W. Liu, “CosFace: Large Margin Cosine Loss for Deep Face Recognition,” 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2018

    2018

  16. [24]

    Partial FC: Training 10 Million Identities on a Single Ma- chine,

    X. An, X. Zhu, Y . Xiao, L. Wu, M. Zhang, Y . Gao, B. Qin, D. Zhang, and Y . Fu, “Partial FC: Training 10 Million Identities on a Single Ma- chine,” 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), 2020

    2021

  17. [25]

    Attention is All you Need,

    A. Vaswani, N. M. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is All you Need,” in NIPS, 2017

    2017

  18. [26]

    Robust Object Recogni- tion Through Symbiotic Deep Learning In Mobile Robots,

    J. Cartucho, R. Ventura, and M. Veloso, “Robust Object Recogni- tion Through Symbiotic Deep Learning In Mobile Robots,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018

    2018

  19. [27]

    The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results,

    M. Everingham, L. Van Gool, C. Williams, J. Winn, and A. Zis- serman, “The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results,” http://www.pascal-network.org/challenges/VOC/ voc2012/workshop/index.html

    2012

  20. [28]

    Technical Report for Ego4D Long Term Action Anticipation Challenge 2023,

    T. Ishibashi, K. Ono, N. Kugo, and Y . Sato, “Technical Report for Ego4D Long Term Action Anticipation Challenge 2023,” 2023. [Online]. Available: https://arxiv.org/abs/2307.01467

    2023 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.