FrameExit

The inference code for FrameExit paper presented in CVPR 2021 (oral paper).

Amir Ghodrati^*,1, Babak Ehteshami Bejnordi^*,1, Amirhossein Habibian¹, "FrameExit: Conditional Early Exiting for Efficient Video Recognition", CVPR 2021 [arxiv].

^* Equal contribution

¹ Qualcomm AI Research (Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc)

Reference

If you find our work useful for your research, please cite:

@inproceedings{ghodrati2021,
  title={FrameExit: Conditional Early Exiting for Efficient Video Recognition},
  author={Ghodrati, Amir and Bejnordi, Babak Ehteshami and Habibian, Amirhossein},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2021}
}

Method

Install

This repository has been tested in Ubuntu 16.04 and Cuda 10. Clone this repository and follow these steps:

conda create -n frameexit python=3.6
conda activate frameexit
conda install pytorch=1.3.1 torchvision=0.4.2 pillow pyyaml
pip install fire==0.3.1 pytorch-ignite==0.3.0

Pretrained models

Our pretrained models can be downloaded from here and should be placed inside the directory resources/checkpoints:

activitynet_checkpoint_25gmac.pth

minikinetics_checkpoint_19.7gmac.pth

Evaluation

To get help:

python3.6 inference.py config/activitynet_inference_2d.yml --help

To get the results on ActivityNet1.3, run the following command:

python3.6 inference.py config/activitynet_inference_2d.yml --data.path_frame <path/to/activitynet/frames> --checkpoint.init <path/to/activitynet/model>

For Mini-Kinetics:

python3.6 inference.py config/minikinetics_inference_2d.yml --data.path_frame <path/to/minikinetics/frames> --checkpoint.init <path/to/minikinetics/model>

where <path/to/.../frames> is the path to the extracted frames, and <path/to/.../model> points to the model file.

Results

After running the above commands, the results should look like:

Name	mAP	Top-1 Acc	GFLOPs
ActivityNet	0.7612	-	24.7
Mini-Kinetics	-	0.7331	19.7

Please note that the results could be slightly different as some videos in the validation set may no longer be available.

In our paper, in addition to Resnet-50, we report the results using the Efficientnet-B3 and X3D-S backbones (Left table). We also report results on the HVU dataset (Right table):

The accuracy vs. efficiency curves on ActivityNet is shown below:

To reproduce the curves for FrameExit (in black), we use the following data points:

Gflops_resnet50_FrameExit = [17.96, 19.89,  23.78, 26.56, 29.09, 41.2]
map_resnet50_FrameExit = [72.72, 73.90, 75.37, 76.05, 76.41, 77.30]

Gflops_efficientnet_FrameExit = [10.03, 10.98, 14.85, 18.00]
map_efficientnet_FrameExit = [78.57, 79.41, 80.85, 81.18]

dreamerlin / FrameExit