Bringing Online Egocentric Action Recognition Into the Wild

To enable a safe and effective human-robot cooperation, it is crucial to develop models for the identification of human activities. Egocentric vision seems be viable solution solve this problem, therefore many works provide deep learning solutions infer actions from first person videos. However, although very promising, most these do not consider major challenges that comes with realistic deployment, such as portability model, need real-time inference, robustness respect novel domains (i.e., new spaces, users, tasks). With paper, we set boundaries egocentric should applications, defining setting action recognition in wild, which encourages researchers novel, applications-aware solutions. We also present model-agnostic technique enables rapid repurposing existing architectures context, demonstrating feasibility deploy model on tiny device (Jetson Nano) perform task directly edge low energy consumption (2.4W average at 50 fps). The code publicly available at: https://github.com/EgocentricVision/EgoWild.