383beaea4aa57dd8202dbff464fee3af-AuthorFeedback.pdf

Ourmodel19 uses unstructured features provided by a convnet backbone and nodes are not associated with specific entities (e.g.20 objects or joints). This has an advantage: it permits independent end-to-end learning and inference, with no need21 externaldetectors. Regarding comparisons to [B], we point out that our main task is activity recognition, whereas [B] tackles action24 localization. Thus, direct comparison is not as trivial. Thus, adapting our graph model to work with an external detector is indeed not36 trivial.