0.25
0.5
0.75
1.25
1.5
1.75
2
Spot On: Action Localization from Pointly-Supervised Proposals
Published on Oct 24, 20161318 Views
We strive for spatio-temporal localization of actions in videos. The state-of-the-art relies on action proposals at test time and selects the best one with a classifier trained on carefully annotated
Related categories
Chapter list
Spot On: Action Localization from Pointly-Supervised Proposals00:00
Goal00:02
Related work: Action proposals at test time/100:16
Related work: Action proposals at test time/200:37
Related work: Training from box supervision00:43
Our hypothesis01:12
Our contribution01:34
Mining the best proposals/101:53
Mining the best proposals/202:20
Proposal affinity/102:45
Proposal affinity/203:21
Proposal match03:29
Size regularization03:57
Mining recap04:21
Experiments04:36
Training without ground truth tubes/105:00
Training without ground truth tubes/205:24
Training without ground truth tubes/305:42
Training without ground truth tubes/405:52
Lowering the annotation frame-rate/106:30
Lowering the annotation frame-rate/207:04
Lowering the annotation frame-rate/307:27
Hollywood2Tubes07:44
Hollywood2Tubes - good example08:22
Hollywood2Tubes - bad example08:44
Hollywood2Tubes - Overall performance09:06
Conclusions 09:23