Tragedy Plus Time: Capturing Unintended Human Activities from Weakly-labeled Videos

Robustness in Sequential Data Workshop @ CVPR 2022

Project Website: https://asu-apg.github.io/TragedyPlusTime/

Dataset: https://drive.google.com/drive/folders/1yy8P208lJdD4dc7KXjA_GUAzimQX9C4n?usp=sharing

Train Model

python main.py --model-name gru --p 1000 --q 10 --Lambda 0.8

Evaluation

Classification Performance

Architecture	Feature	GOAL cMAP	UNINT cMAP
STPN	I3D	45.3	37.5
WTALC	I3D	50.2	38.2
A2CL-PT	I3D	48.5	34.8
Ours	I3D	52.6	41.1

Localization Performance

Model	Feature	Segment	AVG mAP@IoU
STPN	I3D	Goal	21.5
		Unint	15.0
WTALC	I3D	Goal	19.7
		Unint	8.8
A2CL-PT	I3D	Goal	22.2
		Unint	12.5
Ours	I3D	Goal	35.0
		Unint	26.0