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.
Our model discovers sounding actions from in-the-wild egocentric videos. Below we show the top and bottom test examples sorted by our audio-visual similarity scores.
Top examples. Note how the visual activity causes sound in each example.
Bottom exampls. Even though these videos have lots of background sounds correlated with the visual environment, our model successfully assigns them lower scores, accounting for how the sounds are not produced by the action itself.
With our learned embeddings, we can perform crossmodal retrieval between audio, language, and video.
Video-to-audio retrieval
Audio-to-Language retrieval
In this video, we include examples of Ego4D clips, qualitative examples of sounding action discovery, and examples of sounding action retrieval. Wear headphones to hear the sound.