Representation Flow for Action Recognition

CVPR 2019 AJ PiergiovanniMichael S. Ryoo

In this paper, we propose a convolutional layer inspired by optical flow algorithms to learn motion representations. Our representation flow layer is a fully-differentiable layer designed to capture the `flow' of any representation channel within a convolutional neural network for action recognition... (read more)

PDF Abstract

Evaluation Results from the Paper


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK USES EXTRA
TRAINING DATA
COMPARE
Activity Recognition In Videos HMDB51 RepFlow Accuracy 81.1 # 1
Action Classification HMDB51 RepFlow-50 Accuracy 81.1 # 2
Activity Recognition In Videos HMDB-51 RepFlow Accuracy 81.1 # 1
Activity Recognition In Videos HMDB-51 RepFlow Average accuracy of 3 splits 81.1 # 1
Activity Recognition In Videos HMDB-51 RepFlow-50 Average accuracy of 3 splits 81.1 # 1
Action Recognition In Videos HMDB-51 RepFlow-50 Average accuracy of 3 splits 81.1 # 4
Action Classification Kinetics-400 RepFlow-50 Accuracy 77.9 # 4