Adaptive Object Detection

Adaptive approximation for video object detection.

This is the codebase for the paper AdaScale: Towards Real-time Video Object Detection using Adaptive Scaling

Usage

To use the current working version of inference loss.

python experiment/rfcn/rfcn_testloss.py --cfg experiment/rfcn/cfg/rfcn_vid_demo.yaml

Please remember to put ImageNet dataset under /data/imagenet/ILSVRC Also, please put the trained weights under ./output/rfcn/imagenet_vid/rfcn_vid_demo/DET_train_30classes_VID_train_15frames/rfcn_vid-0000.params (What I did is symlink rfcn_vid-0000.params from /data/models/dff_mxnet/rfcn_vid-0000.params)

Citation

If you find this repository useful for your research, please consider citing

@incollection{mlsys2019_209,
author = {Chin, Ting-Wu and Ding, Ruizhou and Marculescu, Diana},
booktitle = {Proceedings of Machine Learning and Systems 2019},
pages = {431--441},
title = {AdaScale: Towards Real-time Video Object Detection using Adaptive Scaling},
year = {2019}
}

Disclaimer

This repository is based on Deep Feature Flow for Video Recognition. We do not own those code written by previous contributers.

The code is not tested and is used for references. The code for AdaScale is under rfcn/