SoundingActions: Learning How Actions Sound from Narrated Egocentric Videos
This addresses the challenge of discovering audio-visual correspondences for a wide range of human actions in egocentric video, which is incremental as it builds on existing multimodal embedding methods.
The paper tackles the problem of learning how actions sound from narrated egocentric videos by proposing a self-supervised embedding method, achieving improved performance over recent techniques on datasets like Ego4D and EPIC-Sounds.
We propose a novel self-supervised embedding to learn how actions sound from narrated in-the-wild egocentric videos. Whereas existing methods rely on curated data with known audio-visual correspondence, our multimodal contrastive-consensus coding (MC3) embedding reinforces the associations between audio, language, and vision when all modality pairs agree, while diminishing those associations when any one pair does not. We show our approach can successfully discover how the long tail of human actions sound from egocentric video, outperforming an array of recent multimodal embedding techniques on two datasets (Ego4D and EPIC-Sounds) and multiple cross-modal tasks.