Audio-Visual Instance Discrimination with Cross-Modal Agreement

CVPR 2021  ·  Pedro Morgado, Nuno Vasconcelos, Ishan Misra ·

We present a self-supervised learning approach to learn audio-visual representations from video and audio. Our method uses contrastive learning for cross-modal discrimination of video from audio and vice-versa. We show that optimizing for cross-modal discrimination, rather than within-modal discrimination, is important to learn good representations from video and audio. With this simple but powerful insight, our method achieves highly competitive performance when finetuned on action recognition tasks. Furthermore, while recent work in contrastive learning defines positive and negative samples as individual instances, we generalize this definition by exploring cross-modal agreement. We group together multiple instances as positives by measuring their similarity in both the video and audio feature spaces. Cross-modal agreement creates better positive and negative sets, which allows us to calibrate visual similarities by seeking within-modal discrimination of positive instances, and achieve significant gains on downstream tasks.

PDF Abstract CVPR 2021 PDF CVPR 2021 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Self-Supervised Audio Classification ESC-50 AVID Top-1 Accuracy 89.2 # 3
Audio Classification ESC-50 AVID Top-1 Accuracy 89.2 # 17
Self-Supervised Action Recognition HMDB51 AVID+CMA (Modified R2+1D-18 on Kinetics) Top-1 Accuracy 60.8 # 25
Pre-Training Dataset Kinetics400 (Video+Audio) # 1
Frozen false # 1
Self-Supervised Action Recognition HMDB51 AVID (Modified R2+1D-18 on Audioset) Top-1 Accuracy 64.1 # 20
Pre-Training Dataset Audioset (Video+Audio) # 1
Frozen false # 1
Self-Supervised Action Recognition HMDB51 AVID+CMA (Modified R2+1D-18 on Audioset) Top-1 Accuracy 64.7 # 16
Pre-Training Dataset Audioset (Video+Audio) # 1
Frozen false # 1
Self-Supervised Action Recognition HMDB51 AVID (Modified R2+1D-18 on Kinetics) Top-1 Accuracy 59.9 # 27
Pre-Training Dataset Kinetics400 (Video+Audio) # 1
Frozen false # 1
Self-Supervised Action Recognition HMDB51 (finetuned) AVID Top-1 Accuracy 64.7 # 8
Self-Supervised Action Recognition UCF101 AVID (Modified R2+1D-18 on Kinetics) 3-fold Accuracy 86.9 # 27
Pre-Training Dataset Kinetics400 (Audio+Video) # 1
Frozen false # 1
Self-Supervised Action Recognition UCF101 AVID+CMA (Modified R2+1D-18 on Kinetics) 3-fold Accuracy 87.5 # 26
Pre-Training Dataset Kinetics400 (Audio+Video) # 1
Frozen false # 1
Self-Supervised Action Recognition UCF101 AVID (Modified R2+1D-18 on Audioset) 3-fold Accuracy 91.0 # 21
Pre-Training Dataset Audioset (Audio+Video) # 1
Frozen false # 1
Self-Supervised Action Recognition UCF101 AVID+CMA (Modified R2+1D-18 on Audioset) 3-fold Accuracy 91.5 # 18
Pre-Training Dataset Audioset (Audio+Video) # 1
Frozen false # 1
Self-Supervised Action Recognition UCF101 (finetuned) AVID 3-fold Accuracy 91.5 # 7

Methods


No methods listed for this paper. Add relevant methods here