Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset

CVPR 2017 Joao CarreiraAndrew Zisserman

The paucity of videos in current action classification datasets (UCF-101 and HMDB-51) has made it difficult to identify good video architectures, as most methods obtain similar performance on existing small-scale benchmarks. This paper re-evaluates state-of-the-art architectures in light of the new Kinetics Human Action Video dataset... (read more)

PDF Abstract

Evaluation Results from the Paper


 SOTA for Action Recognition In Videos on UCF101 (using extra training data)

     Get a GitHub badge
TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK USES EXTRA
TRAINING DATA
COMPARE
Action Classification Charades I3D MAP 32.9 # 6
Hand-Gesture Recognition EgoGesture I3D Accuracy 92.78 # 3
Action Recognition In Videos EgoGesture I3D Accuracy 89.7 # 4
Action Classification HMDB51 Two-stream I3D Accuracy 80.9 # 3
Skeleton Based Action Recognition J-HMDB I3D Accuracy (RGB+pose) 84.1 # 4
Action Classification Moments in Time I3D Top 1 Accuracy 29.51% # 4
Action Classification Moments in Time I3D Top 5 Accuracy 56.06% # 3
Action Recognition In Videos UCF101 Two-stream I3D (on pre-trained) 3-fold Accuracy 98.0 # 1
Action Recognition In Videos UCF101 Two-stream I3D 3-fold Accuracy 93.4 # 14
Action Recognition In Videos VIVA Hand Gestures Dataset I3D Accuracy 83.10 # 2