Abstract
Advances in deep learning have enabled the development of models that have exhibited a remarkable tendency to recognize and even localize actions in videos. However, they tend to experience errors when faced with scenes or examples beyond their initial training environment. Hence, they fail to adapt to new domains without significant retraining with large amounts of annotated data. Current algorithms are trained in an inductive learning environment where they use data-driven models to learn associations between input observations with a fixed set of known classes. In this paper, we propose to overcome these limitations by moving to an open world setting by decoupling the ideas of recognition and reasoning. Building upon the compositional representation offered by Grenander's Pattern Theory formalism, we show that attention and commonsense knowledge can be used to enable the self-supervised discovery of novel actions in egocentric videos in an open-world setting, a considerably more difficult task than zero-shot learning and (un)supervised domain adaptation tasks where target domain data (both labeled and unlabeled) are available during training. We show that our approach can be used to infer and learn novel classes for open vocabulary classification in egocentric videos and novel object detection with zero supervision. Extensive experiments show that it performs competitively with fully supervised baselines on publicly available datasets under open-world conditions. This is one of the first works to address the problem of open-world action recognition in egocentric videos with zero human supervision to the best of our knowledge.
Abstract (translated)
URL
https://arxiv.org/abs/2009.07470
