The application of visual-audio fixation prediction. A. Kinematics, B. Criminal Psychology, C. Airplane Pilots’ Skill Training, D. Lie Detection, E. Operative Risk Assessment, F. 360 Video Surveillance.

• This repository contains the source code, results, and evaluation toolbox of STANet+ (V2), which are the journal extension version of our paper STANet (V₁) published at CVPR-2021.
• Compared our conference version STANet (V₂), which has been extended in two distinct aspects.
  First on the basis of multisource and multiscale perspectives which have been adopted by the CVPR version (V1), we have provided a deep insight into the relationship between multigranularity perception (Fig.2) and real human attention behaved in visual-auditory environment.
  Second without using any complex networks, we have provided an elegant framework to complementary integrate multisource, multiscale, and multigranular information (Fig.1) to formulate pseudofixations which are very consistent with the real ones. Apart from achieving significant performance gain, this work also provides a comprehensive solution for mimicking multimodality attention.

Figure 1: STANet+ mainly focuses on devising a weakly supervised approach for the spatial-temporal-audio (STA) fixation prediction task, where the key innovation is that, as one of the first attempts, we automatically convert semantic category tags to pseudofixations via the newly proposed selective class activation mapping (SCAM) and the upgraded version SCAM+ that has been additionally equipped with the multigranularity perception ability. The obtained pseudofixations can be used as the learning objective to guide knowledge distillation to teach two individual fixation prediction networks (i.e., STA and STA+), which jointly enable generic video fixation prediction without requiring any video tags.

Figure 2: Some representative ’fixation shifting’ cases, additional multigranularity information (i.e., long/crossterm information) has been shown before collecting fixations in A_SRC. Clearly, by comparing A_FIX0, A_FIX1, and A _FIX2, we can easily notice that the multigranularity information could draw human attention to the most meaningful objects and make the fixations to be more focused.

• Windows10
• NVIDIA GeForce RTX 2070 SUPER & NVIDIA GeForce RTX 1080Ti
• python 3.6.4
• Matlab R2016b
• pytorch 1.8.0
• soundmodel
• Linux
• TITAN V * 8

https://pan.baidu.com/s/12hiunJrdi5fNjftZsVRTKQ Code:r5fs

Visual:resnext101_32x8d, vgg16
Audio: vggsound, net = torch.load('vggsound_netvlad').

Training dataset: AVE(Audio Visual Event Location).
Testing dataset: AVAD, DIEM, SumMe, ETMD, Coutrot.

Note
We use Fourier-transform to transform audio features as audio stream input, therefore, you firstly need to use the function audiostft.py to convert the audio files (.wav) to get the audio features(.h5).

Coarse: Separately training branches of S_coarse, SA_coarse, ST_coarse ，it should be noted that the coarse stage is coarse location, so the size is set to 256 to ensure object-wise location accuracy.
Fine: Separately re-training branches of S_fine, SA_fine, ST_fine，it should be noted that the fine stage is a fine location, so the size is set to 356 to ensure regional location exactness.

S+: Separately training branches of S+_short, S+_long, S+_cross, because it is frame-wise relational reasoning network, the network is the same, so we only need to change the source of the input data.
SA+: Separately training branches of SA+_long, SA+_cross.
ST+: Separately training branches of ST+_short, ST+_long, ST+_cross.

In order to facilitate the display of matrix data processing, Matlab2016b was performed in coarse location of inter-frame smoothing and pseudo GT data post-processing.

Training the model of STA and STA+ using the AVE video frames with the generated pseudoGT.

The model weight file STANet+, STANet, AudioSwitch:
(Baidu Netdisk, code:r5fs).

(STANet+, STANet, ITTI, GBVS, SCLI, AWS-D, SBF, CAM, GradCAM, GradCAMpp, SGradCAMpp, xGradCAM, SSCAM, ScoCAM, LCAM, ISCAM, ACAM, EGradCAM, ECAM, SPG, VUNP, WSS, MWS, WSSA).
(Baidu Netdisk, code:6afo).

Qualitative results of our method and eight representative saliency models: ITTI, GBVS, SCLI, SBF, AWS-D, WSS, MWS, WSSA. It can be observed that our method is able to handle various challenging scenes well and produces more accurate results than other competitors.

Quantitative comparisons between our method with other fully-/weakly-/un-supervised methods on 6 datasets. Bold means the best result, " denotes the higher the score, the better the performance.

[1][Tsiami, A., Koutras, P., Maragos, P.STAViS: Spatio-Temporal AudioVisual Saliency Network. (CVPR 2020).] (https://openaccess.thecvf.com/content_CVPR_2020/papers/Tsiami_STAViS_Spatio-Temporal_AudioVisual_Saliency_Network_CVPR_2020_paper.pdf)
[2][Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C. Audio-Visual Event Localization in Unconstrained Videos. (ECCV 2018)] (https://openaccess.thecvf.com/content_ECCV_2018/papers/Yapeng_Tian_Audio-Visual_Event_Localization_ECCV_2018_paper.pdf)
[3][Chen, H., Xie, W., Vedaldi, A., & Zisserman, A. Vggsound: A Large-Scale Audio-Visual Dataset. (ICASSP 2020)] (https://www.robots.ox.ac.uk/~vgg/publications/2020/Chen20/chen20.pdf)

If you find this work useful for your research, please consider citing the following paper:

@InProceedings{Wang_2021_CVPR,  
    author    = {Wang, Guotao and Chen, Chenglizhao and Fan, Deng-Ping and Hao, Aimin and Qin, Hong},
    title     = {From Semantic Categories to Fixations: A Novel Weakly-Supervised Visual-Auditory Saliency Detection Approach},  
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},  
    month     = {June},  
    year      = {2021},  
    pages     = {15119-15128}  
}  


@article{wang2021weakly,
   title={Weakly Supervised Visual-Auditory Fixation Prediction with Multigranularity Perception},
   author={Wang, Guotao and Chen, Chenglizhao and Fan, Deng-ping and Hao, Aimin and Qin, Hong},
   journal={arXiv preprint arXiv:2112.13697},
   year={2021}
}