Abductive Action Inference

Abductive reasoning aims to make the most likely inference for a given set of incomplete observations. In this work, we propose a new task called abductive action inference, in which given a situation, the model answers the question `what actions were executed by the human in order to arrive in the current state?'. Given a state, we investigate three abductive inference problems: action set prediction, action sequence prediction, and abductive action verification. We benchmark several SOTA models such as Transformers, Graph neural networks, CLIP, BLIP, end-to-end trained Slow-Fast, and Resnet50-3D models. Our newly proposed object-relational BiGED model outperforms all other methods on this challenging task on the Action Genome dataset. Codes will be made available.