Pith. sign in

REVIEW 4 major objections 4 minor 41 references

Minimalistic Video Saliency Prediction via Efficient Decoder & Spatio Temporal Action Cues

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read The paper claims that an ensemble of two compact 3D convolutional models—one with a lightweight grouped-convolution decoder, one with action-localization features—sets state-of-the-art video saliency results on nine datasets while running…

desk verdict The efficiency story for ViNet-S and the ensemble numbers hold up, but the paper's causal claim about STAL features is confounded and the abstract overclaims SOTA. read the letter →

arxiv 2502.00397 v1 pith:SMGIBDEO submitted 2025-02-01 cs.CV

classification cs.CV
keywords videosaliencypredictionefficientdeeplearningspatio-temporalactionlocalizationgroupedconvolutionschannelshuffleSlowFastensembleaveragingaudio-visual
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 sets out to show that video saliency prediction does not require giant transformer stacks or audio channels: a small U-Net-style 3D convolutional model with a decoder built from grouped convolutions and channel shuffles matches or beats much larger models, and averaging it with a variant that uses spatio-temporal action localization features yields state-of-the-art results on three visual-only and six audio-visual datasets. The central claim is that the gain comes from using an action localization backbone rather than an action classification backbone, because the localization features capture interactions between actors and their context that classification features miss. If true, video saliency models can be dramatically smaller and faster without sacrificing accuracy, and audio information is largely unnecessary for the tested benchmarks.

What carries the argument

The mechanism is a 3D convolutional encoder-decoder whose decoder uses filter groups followed by channel shuffle layers to cut parameters, paired with two different pretrained video encoders. ViNet-S uses the S3D action-classification backbone; ViNet-A uses the SlowFast two-pathway network pretrained on AVA, a spatio-temporal action localization dataset, with a neck that fuses the slow and fast pathways before decoding. The decoder itself contains only 1.6 million parameters, so the backbone dominates ViNet-A's size. The ensemble works by averaging the two saliency maps, which the paper interprets as combining complementary action-classification-style and action-localization-style cues.

What would settle it

Train ViNet-A's SlowFast backbone on Kinetics action classification instead of AVA action localization, keeping the decoder and training procedure identical, and compare on human-centric datasets such as MVVA and Coutrot2; if the Kinetics-pretrained version matches the AVA-pretrained version, the paper's central attribution to localization cues fails.

Watch

Extended reading notes

Core claim

On its own terms, this paper claims that a deliberately minimal visual-only architecture can outperform audio-visual transformer-based models on video saliency. ViNet-S, the ViNet encoder with a lightweight decoder using grouped convolutions and channel shuffles, shrinks the model to 36MB and 9.5 million parameters while consistently beating the original 124MB ViNet. ViNet-A swaps the encoder for a SlowFast network pretrained for spatio-temporal action localization on AVA, which the paper credits with markedly better performance on human-centric datasets. ViNet-E, the pixel-wise average of the two predicted saliency maps, achieves top results across the nine datasets, and the ensemble remains smaller in size and parameters than transformer-based alternatives while running in real time.

Load-bearing premise

The claim that localization features rather than classification features drive the gains rests on comparing models that differ in two ways at once—backbone architecture and pretraining data—so the contribution of the localization cue alone is never isolated.

Editorial extensions

If this is right

  • A 36MB, 9.5-million-parameter model can recover at least 98.5% of a transformer-based model's correlation score on Hollywood2 while using about six times fewer parameters and beating it on AUC-Judd.
  • ViNet-A improves normalized scanpath saliency by over 20% relative to a multi-branch audio-visual model on the MVVA dataset while using only the visual modality.
  • Real-time deployment is feasible: ViNet-S runs at about 200fps, ViNet-A at 120fps, and the ensemble at 90fps at batch size one, with ViNet-S reaching 1070fps at batch size eight on an RTX 4090.
  • Because pixel-wise averaging of the two models consistently helps, action classification features and action localization features appear to carry complementary information for saliency.
  • Several audio-visual baselines behave identically when given random or zero audio at inference, so their reported audio-driven gains should be treated with caution.

Reading between the lines

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

  • The paper leaves the backbone attribution untested: ViNet-A differs from ViNet-S in both architecture and pretraining data, so directly replacing ViNet-A's AVA pretraining with Kinetics pretraining on the same SlowFast backbone would isolate whether localization cues or simply model capacity drive the gains.
  • The success of simple map averaging suggests a cheap extension: pair ViNet-S with any other localization-based saliency model and average their outputs without retraining, potentially transferring the ensemble benefit to other backbones.
  • The audio-agnostic finding implies that future audio-visual saliency benchmarks should include audio-stripping controls as a standard check before crediting a model with true multimodal understanding.
  • If the STAL advantage holds, an even stronger model might come from fine-tuning the action-localization backbone on saliency data with multi-person annotations, directly encouraging the detector to focus on salient faces and interactions.
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 / 4 minor

Summary. The paper proposes two video saliency models, ViNet-S and ViNet-A, built on a common lightweight decoder with grouped convolutions and channel shuffles. ViNet-S keeps the S3D backbone from ViNet, while ViNet-A replaces it with a SlowFast backbone pretrained on AVA for spatio-temporal action localization. A third model, ViNet-E, averages the predicted saliency maps of ViNet-S and ViNet-A. The authors evaluate on three visual-only datasets (DHF1K, Hollywood2, UCF-Sports) and six audio-visual datasets (AVAD, Coutrot1, Coutrot2, DIEM, ETMD, MVVA), reporting four standard metrics plus model size, parameter count, and inference speed. They claim state-of-the-art performance on all nine datasets, with ViNet-S reaching over 1000 fps.

Significance. If the reported results hold, the contribution is a compact, fast, visual-only saliency model family that is competitive with much larger transformer-based methods, and an interesting hypothesis that spatio-temporal action localization features help in human-centric scenes. The breadth of the evaluation (nine datasets, multiple metrics) and the parameter-efficiency analysis are strengths. However, the central causal claim about the STAL backbone is not yet supported by controlled experiments, and the headline state-of-the-art statement is internally contradicted by the paper's own tables. These issues need to be addressed before the contribution is fully substantiated.

major comments (4)
  1. [IV-A, Tables I(a), II, III] The sentence in Section IV-A that ViNet-A's strong performance on UCF-Sports and Hollywood2 'clearly demonstrates the advantages of using an STAL backbone over an action classification backbone' is not supported by the presented comparisons. ViNet-A and ViNet-S differ in three ways at once: backbone architecture (SlowFast vs S3D), pretraining data and task (AVA action localization vs Kinetics action classification), and input sampling (32 frames sampled from a 64-frame window with supervision on the 33rd frame vs 32 consecutive frames with supervision on the 32nd frame). Any of these differences, or their interaction, could explain the gains on human-centric datasets. Please add a controlled ablation, e.g., a SlowFast backbone pretrained on Kinetics or an S3D backbone pretrained on AVA, keeping the decoder, loss, and input sampling fixed; without such an ablation, contribution (2) is not established.
  2. [Abstract and Tables I(a), II] The abstract's claim that the ensemble ViNet-E achieves state-of-the-art performance on three visual-only and six audio-visual datasets is contradicted by the paper's own numerical results. On DHF1K (Table I(a)), ViNet-E has CC 0.549, NSS 3.134, AUC-J 0.922, and SIM 0.409, all below TMFI-Net's 0.554, 3.201, 0.924, and 0.428. On Coutrot2 (Table II), ViNet-A achieves CC 0.862 while ViNet-E achieves 0.854, so the ensemble is not the best method on that dataset either. Please revise the claim to reflect these exceptions, or state explicitly the per-dataset, per-metric basis for the state-of-the-art claim.
  3. [IV-d and III-b] The 'over 1000fps' statement is reported only for batch size eight, whereas the real-time single-stream setup gives approximately 200fps for ViNet-S; the paper should clearly separate batched throughput from single-stream latency and ideally report FLOPs. More importantly, the paper reports no error bars or significance tests, even though several datasets use 3-fold cross-validation and MVVA uses a single random split. Differences on the order of 0.01 in CC may be within run-to-run or split-to-split variability, so the headline ranking claims would be more convincing with per-split results or confidence intervals.
  4. [IV-b, Tables II and III] It is unclear which of the baseline rows in Tables II and III correspond to audio-visual models and which to their visual-only versions. The text says 'we limit our comparisons only to their visual only model,' but the tables list methods such as STA ViS, TSFP-Net, and CASP-Net, which are originally audio-visual. Please clarify in the tables or captions which results are re-runs of visual-only variants and which are taken from the original audio-visual papers, since the claim of surpassing audio-visual models without using audio depends on this distinction.
minor comments (4)
  1. [Table I(b)] The table row uses 'STSA-Net' while the text uses 'STSANet'; please unify the name of this method.
  2. [Table II] The dataset header 'A V AD' contains extra spaces; use 'AVAD' consistently throughout the paper.
  3. [III-b] The MVVA random split is not specified (for example, training/test proportions or random seed); please report the split details to improve reproducibility.
  4. [IV-d] The paper reports runtimes only for an RTX 4090; adding hardware details and measured frame rates for a common input resolution would make the real-time comparison more reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the paper's claims are empirical benchmark results supported by external datasets and prior published baselines, with no derivation reducing to its own inputs.

full rationale

ViNet-S/A/E is an empirical benchmark paper rather than a derivation. The central claims — state-of-the-art results on nine saliency datasets, parameter efficiency, and real-time inference speeds — are supported by evaluations on external datasets and comparisons to previously published methods. The loss function (KLDiv − CC) and decoder filter-group counts were selected by experimentation on the same benchmarks, which is a mild selection-on-test-data risk, but not a circular reduction: the reported metrics are still measured on held-out splits against external baselines. The paper cites its own prior ViNet [12] and its own prior audio-study [41]; these citations provide an architectural starting point and a supporting prior finding, respectively, and are not invoked to forbid alternatives or to justify the central SOTA claim. The strongest non-circular weakness is that the ViNet-A versus ViNet-S comparison conflates backbone family, pretraining task/data, and input sampling, making the causal attribution to spatio-temporal action localization under-supported; that is a confound and a correctness risk, not circularity. No step reduces to its own input by construction.

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

The paper introduces no new physical or mathematical entities. Its central claims rest on three assumptions: the acceptance of standard saliency metrics, the transferability of pretrained video backbones, and the comparability of baseline numbers across different papers. The main free choices are the decoder filter-group layout, the loss function, and the ViNet-A input sampling, none of which is ablated.

free parameters (3)
  • Decoder filter group configuration = 32, 16, 8, 8, 4, 2
    The number of groups for each 3D convolution in the decoder was chosen by experiment: 'We experimented with different filter groups and channel shuffle layer configurations and found this setup optimal.' The reported accuracy depends on this choice.
  • Loss function = KLDiv(P,Q) - CC(P,Q)
    The loss is a weighted-free combination of KL divergence and negative correlation, selected after experimentation across combinations: 'Through experimentation with different combinations, we found that the optimal results for most datasets were achieved with...' This directly affects the trained model's performance.
  • ViNet-A input sampling = 32 frames sampled every other frame from a 64-frame window
    The sampling strategy is adopted 'akin to action label predictions in STAL models' and is not ablated, so its contribution to the ViNet-A result is unknown.
assumptions (3)
  • domain assumption The standard saliency metrics (CC, NSS, AUC-J, SIM, KLDiv) are valid and sufficient for comparing models.
    The paper uses these metrics without justification, following [40], and does not consider whether a single metric would change rankings.
  • domain assumption Pretrained video backbones transfer their features to saliency prediction.
    S3D pretrained on Kinetics and SlowFast pretrained on AVA are used as encoders without evidence that their features correlate with human gaze.
  • domain assumption The numbers reported for baseline methods in their original papers are directly comparable to the authors' reimplementations and training protocols.
    The paper copies baseline results from prior papers and runs new experiments for its own models under different hardware, code, and possibly data splits, assuming comparability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Minimalistic Video Saliency Prediction via Efficient Decoder & Spatio Temporal Action Cues." pith.science (2026). https://pith.science/paper/SMGIBDEO

@misc{pith2026250200397,
  author       = {Pith},
  title        = {Pith review of: Minimalistic Video Saliency Prediction via Efficient Decoder & Spatio Temporal Action Cues},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SMGIBDEO}},
  note         = {Machine review of arXiv:2502.00397}
}
read the original abstract

This paper introduces ViNet-S, a 36MB model based on the ViNet architecture with a U-Net design, featuring a lightweight decoder that significantly reduces model size and parameters without compromising performance. Additionally, ViNet-A (148MB) incorporates spatio-temporal action localization (STAL) features, differing from traditional video saliency models that use action classification backbones. Our studies show that an ensemble of ViNet-S and ViNet-A, by averaging predicted saliency maps, achieves state-of-the-art performance on three visual-only and six audio-visual saliency datasets, outperforming transformer-based models in both parameter efficiency and real-time performance, with ViNet-S reaching over 1000fps.

Figures

Figures reproduced from arXiv: 2502.00397 by the authors.

Figure 1
Figure 1. Our Model (ViNet-A) Architecture for SP (Best viewed in colour) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Qualitative results: Comparing Ground Truth with the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 38 canonical work pages

  1. [1]

    Visual saliency model for robot cameras,

    N. J. Butko, L. Zhang, G. W. Cottrell, and J. R. Movellan, “Visual saliency model for robot cameras,” in 2008 IEEE International Confer- ence on Robotics and Automation . IEEE, 2008, pp. 2398–2403

  2. [2]

    Gazed– gaze-guided cinematic editing of wide-angle monocular video record- ings,

    K. B. Moorthy, M. Kumar, R. Subramanian, and V . Gandhi, “Gazed– gaze-guided cinematic editing of wide-angle monocular video record- ings,” in ACM Conference on Human Factors in Computing Systems (CHI), 2020, pp. 1–11

  3. [3]

    Salgaze: Personalizing gaze estimation using visual saliency,

    Z. Chang, J. Matias Di Martino, Q. Qiu, S. Espinosa, and G. Sapiro, “Salgaze: Personalizing gaze estimation using visual saliency,” in Inter- national Conference on Computer Vision Workshops (ICCVW) , 2019

  4. [4]

    Attentional mechanisms for socially in- teractive robots–a survey,

    J. F. Ferreira and J. Dias, “Attentional mechanisms for socially in- teractive robots–a survey,” IEEE Transactions on Autonomous Mental Development, vol. 6, no. 2, pp. 110–125, 2014

  5. [5]

    Facial expression recognition using visual saliency and deep learning,

    V . Mavani, S. Raman, and K. P. Miyapuram, “Facial expression recognition using visual saliency and deep learning,” in International Conference on Computer Vision Workshops (ICCVW) , 2017

  6. [6]

    Evaluating the effect of saliency detection and attention manipulation in human-robot interac- tion,

    G. Schillaci, S. Bodiro ˇza, and V . V . Hafner, “Evaluating the effect of saliency detection and attention manipulation in human-robot interac- tion,” International Journal of Social Robotics , vol. 5, pp. 139–152, 2013

  7. [7]

    Saliency heat-map as visual attention for autonomous driving using generative adversarial network (gan),

    F. Lateef, M. Kas, and Y . Ruichek, “Saliency heat-map as visual attention for autonomous driving using generative adversarial network (gan),” IEEE Transactions on Intelligent Transportation Systems , vol. 23, no. 6, pp. 5360–5373, 2021

  8. [8]

    A gated fusion network for dynamic saliency prediction,

    A. Kocak, E. Erdem, and A. Erdem, “A gated fusion network for dynamic saliency prediction,” IEEE Transactions on Cognitive and Developmental Systems , vol. 14, no. 3, pp. 995–1008, 2021

Show all 41 references
  1. [9]

    Video saliency prediction based on spatial- temporal two-stream network,

    K. Zhang and Z. Chen, “Video saliency prediction based on spatial- temporal two-stream network,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 29, no. 12, pp. 3544–3557, 2019

  2. [10]

    Unified image and video saliency modeling,

    R. Droste, J. Jiao, and J. A. Noble, “Unified image and video saliency modeling,” in European Conference on Computer Vision (ECCV) . Springer, 2020, pp. 419–435

  3. [11]

    Revisiting video saliency prediction in the deep learning era,

    W. Wang, J. Shen, J. Xie, M.-M. Cheng, H. Ling, and A. Borji, “Revisiting video saliency prediction in the deep learning era,” TPAMI, vol. 43, no. 1, pp. 220–237, 2019

  4. [12]

    Vinet: Pushing the limits of visual modality for audio-visual saliency prediction,

    S. Jain, P. Yarlagadda, S. Jyoti, S. Karthik, R. Subramanian, and V . Gandhi, “Vinet: Pushing the limits of visual modality for audio-visual saliency prediction,” in IROS. IEEE, 2021, pp. 3520–3527

  5. [13]

    Tased-net: Temporally-aggregating spatial encoder-decoder network for video saliency detection,

    K. Min and J. J. Corso, “Tased-net: Temporally-aggregating spatial encoder-decoder network for video saliency detection,” in International Conference on Computer Vision (ICCV) , 2019, pp. 2394–2403

  6. [14]

    Rethinking spatiotem- poral feature learning: Speed-accuracy trade-offs in video classification,

    S. Xie, C. Sun, J. Huang, Z. Tu, and K. Murphy, “Rethinking spatiotem- poral feature learning: Speed-accuracy trade-offs in video classification,” in European Conference on Computer Vision (ECCV) , 2018

  7. [15]

    The kinetics human action video dataset,

    W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijaya- narasimhan, F. Viola, T. Green, T. Back, P. Natsev et al. , “The kinetics human action video dataset,” arXiv preprint arXiv:1705.06950 , 2017

  8. [16]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in MICCAI, 2015

  9. [17]

    Spatio-temporal self-attention network for video saliency prediction,

    Z. Wang, Z. Liu, G. Li, Y . Wang, T. Zhang, L. Xu, and J. Wang, “Spatio-temporal self-attention network for video saliency prediction,” IEEE Transactions on Multimedia , vol. 25, pp. 1161–1174, 2021

  10. [18]

    Transformer-based multi-scale feature integration network for video saliency prediction,

    X. Zhou, S. Wu, R. Shi, B. Zheng, S. Wang, H. Yin, J. Zhang, and C. Yan, “Transformer-based multi-scale feature integration network for video saliency prediction,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 12, pp. 7696–7707, 2023

  11. [19]

    Transformer-based video saliency prediction with high temporal dimension decoding,

    M. Moradi, S. Palazzo, and C. Spampinato, “Transformer-based video saliency prediction with high temporal dimension decoding,” VISI- GRAPP, 2024

  12. [20]

    Stavis: Spatio-temporal audio- visual saliency network,

    A. Tsiami, P. Koutras, and P. Maragos, “Stavis: Spatio-temporal audio- visual saliency network,” in Conference on Computer Vision and Pattern Recognition (CVPR) , 2020, pp. 4766–4776

  13. [21]

    Temporal-spatial feature pyramid for video saliency detection,

    Q. Chang and S. Zhu, “Temporal-spatial feature pyramid for video saliency detection,” arXiv preprint arXiv:2105.04213 , 2021

  14. [22]

    Joint learning of audio-visual saliency prediction and sound source localization on multi-face videos,

    M. Qiao, Y . Liu, M. Xu, X. Deng, B. Li, W. Hu, and A. Borji, “Joint learning of audio-visual saliency prediction and sound source localization on multi-face videos,” International Journal of Computer Vision (IJCV), vol. 132, pp. 2003–2025, 2023

  15. [23]

    Learning to predict salient faces: A novel visual-audio saliency model,

    Y . Liu, M. Qiao, M. Xu, B. Li, W. Hu, and A. Borji, “Learning to predict salient faces: A novel visual-audio saliency model,” in European Conference on Computer Vision (ECCV) , 2020, pp. 413–429

  16. [24]

    Casp-net: Rethinking video saliency prediction from an audio-visual consistency perceptual perspective,

    J. Xiong, G. Wang, P. Zhang, W. Huang, Y . Zha, and G. Zhai, “Casp-net: Rethinking video saliency prediction from an audio-visual consistency perceptual perspective,” in Conference on Computer Vision and Pattern Recognition (CVPR) , 2023, pp. 6441–6450

  17. [25]

    Diffsal: Joint audio and video learning for diffusion saliency prediction,

    J. Xiong, P. Zhang, T. You, C. Li, W. Huang, and Y . Zha, “Diffsal: Joint audio and video learning for diffusion saliency prediction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 27 273–27 283

  18. [26]

    Deep roots: Improving cnn efficiency with hierarchical filter groups,

    Y . Ioannou, D. Robertson, R. Cipolla, and A. Criminisi, “Deep roots: Improving cnn efficiency with hierarchical filter groups,” in Conference on Computer Vision and Pattern Recognition (CVPR) , 2017

  19. [27]

    Shufflenet: An extremely efficient convolutional neural network for mobile devices,

    X. Zhang, X. Zhou, M. Lin, and J. Sun, “Shufflenet: An extremely efficient convolutional neural network for mobile devices,” inConference on Computer Vision and Pattern Recognition (CVPR) , 2018

  20. [28]

    Actor- context-actor relation network for spatio-temporal action localization,

    J. Pan, S. Chen, M. Z. Shou, Y . Liu, J. Shao, and H. Li, “Actor- context-actor relation network for spatio-temporal action localization,” in Conference on Computer Vision and Pattern Recognition (CVPR) , 2021, pp. 464–474

  21. [29]

    Slowfast networks for video recognition,

    C. Feichtenhofer, H. Fan, J. Malik, and K. He, “Slowfast networks for video recognition,” in International Conference on Computer Vision (ICCV), 2019, pp. 6202–6211

  22. [30]

    Ava: A video dataset of spatio-temporally localized atomic visual actions,

    C. Gu, C. Sun, D. A. Ross, C. V ondrick, C. Pantofaru, Y . Li, S. Vi- jayanarasimhan, G. Toderici, S. Ricco, R. Sukthankar, C. Schmid, and J. Malik, “Ava: A video dataset of spatio-temporally localized atomic visual actions,” in Conference on Computer Vision and Pattern Recogn...

  23. [31]

    Actions in the eye: Dynamic gaze datasets and learnt saliency models for visual recognition,

    S. Mathe and C. Sminchisescu, “Actions in the eye: Dynamic gaze datasets and learnt saliency models for visual recognition,” TPAMI, vol. 37, no. 7, pp. 1408–1424, 2014

  24. [32]

    Fixation prediction through multimodal analysis,

    X. Min, G. Zhai, K. Gu, and X. Yang, “Fixation prediction through multimodal analysis,” ACM Trans. Multimedia Comput. Commun. Appl. , vol. 13, no. 1, 2016

  25. [33]

    How saliency, faces, and sound influence gaze in dynamic social scenes,

    A. Coutrot and N. Guyader, “How saliency, faces, and sound influence gaze in dynamic social scenes,” Journal of vision , vol. 14, no. 8, pp. 5–5, 2014

  26. [34]

    Toward the introduction of auditory information in dynamic visual attention models,

    ——, “Toward the introduction of auditory information in dynamic visual attention models,” in 2013 14th International Workshop on Image Analysis for Multimedia Interactive Services (WIAMIS) , 2013, pp. 1–4

  27. [35]

    An efficient audiovisual saliency model to predict eye positions when looking at conversations,

    ——, “An efficient audiovisual saliency model to predict eye positions when looking at conversations,” in 2015 23rd European Signal Process- ing Conference (EUSIPCO) . IEEE, 2015, pp. 1531–1535

  28. [36]

    Clustering of gaze during dynamic scene viewing is predicted by motion,

    P. K. Mital, T. J. Smith, R. L. Hill, and J. M. Henderson, “Clustering of gaze during dynamic scene viewing is predicted by motion,” Cognitive Computation, vol. 3, pp. 5–24, 2011

  29. [37]

    Predicting eyes’ fixa- tions in movie videos: Visual saliency experiments on a new eye- tracking database,

    P. Koutras, A. Katsamanis, and P. Maragos, “Predicting eyes’ fixa- tions in movie videos: Visual saliency experiments on a new eye- tracking database,” in Engineering Psychology and Cognitive Er- gonomics, D. Harris, Ed., 2014, pp. 183–194

  30. [38]

    Tinyhd: Efficient video saliency prediction with heterogeneous decoders using hierarchical maps distillation,

    F. Hu, S. Palazzo, F. P. Salanitri, G. Bellitto, M. Moradi, C. Spampinato, and K. McGuinness, “Tinyhd: Efficient video saliency prediction with heterogeneous decoders using hierarchical maps distillation,” in Winter Conference on Applications of Computer Vision (WACV) , 2023

  31. [39]

    Video saliency forecasting transformer,

    C. Ma, H. Sun, Y . Rao, J. Zhou, and J. Lu, “Video saliency forecasting transformer,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 10, pp. 6850–6862, 2022

  32. [40]

    What do different evaluation metrics tell us about saliency models?

    Z. Bylinskii, T. Judd, A. Oliva, A. Torralba, and F. Durand, “What do different evaluation metrics tell us about saliency models?” vol. 41, no. 3, pp. 740–757, 2019

  33. [41]

    Does audio help in deep audio-visual saliency prediction models?

    R. Agrawal, S. Jyoti, R. Girmaji, S. Sivaprasad, and V . Gandhi, “Does audio help in deep audio-visual saliency prediction models?” in Proceed- ings of the 2022 International Conference on Multimodal Interaction , 2022, pp. 48–56

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.