Facebook AI aims to step towards egocentric perception and, for this purpose, they recently released the largest first-person perspective dataset currently available, EGO4D, opens an external URL in a new window. The EGO4D Action Anticipation Challenge, opens an external URL in a new window aimed at answering "What am I likely to do next?" in the long-term future, based on the video observation of the last actions performed by the human.

Our team, formed by Esteve Valls Mascaro and Prof. Dongheui Lee, from TUWien Autonomous System Labs, and Prof Hyemin Ahn, from Ulsan National Institute (UNIST), presented our work in the EGO4D Workshop in CVPR@2022, opens an external URL in a new window in New Orleans last June 19th, 2022. Our framework first extracts the human intention from the past observation of the person to anticipate a more realistic sequence of actions that this human will execute in the future. We call our method Intention-Conditioned Variational Autoencoder (I-CVAE).

We are happy to claim that our proposed work is able to advance the exploration of human behavior and excel in predicting the next actions of this human compared to the baseline, with direct applications in task planning and human-robot collaborative scenarios. We plan to release a publication in the near future with more details of our approach. Let's keep in touch!