TengdaHan/ActionClassification

Video action classification benchmark for common CNN architectures, implemented in PyTorch

Action Classification

What's this?

A benchmark of video action classification for common CNN architectures. Implemented in PyTorch v1.2, recent version should also work.

Supported networks

Network	#parameters (exclude final classifier)
2d3d-ResNet18	31.82 M
2d3d-ResNet34	60.80 M
2d3d-ResNet50	44.74 M
3d-ResNet18	33.15 M
3d-ResNet34	63.46 M
3d-ResNet50	46.14 M
I3D	12.29 M
S3D	7.910 M
S3D-G	9.098 M

Files

• backbone/ has all backbone models
• model.py gives an example of classifier with S3D backbone.

Notes

• benchmarks will come soon

Link

• ResNet-2d3d is used in SlowFast, DPC, etc.
• ResNet-3d is used in many papers, early ones like Hara et al.
• I3D from Carreira and Zisserman
• S3D/S3D-G from Xie et al.