Moving in Sync: Self-supervised learning of n-human interactions
summary
Introduces Sync-3D, a self‑supervised pipeline that feeds two tracked human crops through an I3D+SyncNet style siamese network trained with contrastive loss and curriculum sampling of easy/medium/hard negatives (hard = same video shifted ~25 frames, positives require temporal alignment and IoU > 0.1). Trained and evaluated on TV‑HID (300 clips, 5 classes) it gets low binary sync accuracy (50.4 train, 42.6 val) but improves downstream interaction classification to 75.5% versus I3D’s 73.2%, with per‑class highs 89% (kiss) and 87% (hug).
AI-written summary · about
chapters (6)
related talks
35mDeep learning for 3d human motion capture
06mLearning When Agents Can Talk to Drivers using the INAGT Dataset and Multisensor Fusion
59mDeep Neural Networks for Video Applications
41mSocial Behavior Prediction from Video Observations
26mComputer Vision in the Era of Deep Learning
06mSyncWISE: Window Induced Shift Estimation for Synchronization ofVideo and Accelerometry from ...
47mDeep Learning in Computer Vision
27mAnalysis of sports scenes using computer vision models