Collaborative Object Detectors Adaptive to Bandwidth and Computation

!News: This paper has been accepted to ICASSP 2022

The final version will be uploaded soon, until then check our preprint @ https://arxiv.org/abs/2105.00591. If you want to cite us please use the bibtex bellow:

@inproceedings{assine2022single,
  title={Collaborative Object Detectors Adaptive to Bandwidth and Computation},
  author={Assine, Juliano S and Santos Filho, JCS and Valle, Eduardo},
  booktitle={ICASSP 2022-2022 IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={4467--4471},
  year={2022},
  organization={IEEE}
}

Also, this repo is based on yoshitomo-matsubara/hnd-ghnd-object-detectors. Go check out their work.

Summary

In this work we explore split object detectors such as the one bellow

Repository setup

How to clone

git clone https://github.com/jsiloto/adaptive-cod
cd adaptive-cod
git submodule init
git submodule update --recursive --remote

COCO 2017 Dataset

Make sure you have the COCO2017 dataset downloaded

mkdir -p ./resource/dataset/coco2017
cd ./resource/dataset/coco2017
wget http://images.cocodataset.org/zips/train2017.zip
wget http://images.cocodataset.org/zips/val2017.zip
wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
unzip -q train2017.zip
unzip -q val2017.zip
unzip -q annotations_trainval2017.zip

Docker environment

We highly recommend using the provided Dockerfile for setting up your environment.

docker build -f Dockerfile -t adaptive_cod .

docker run --gpus all --rm -it \
 -p 8888:8888 -v $PWD:/work --shm-size=16G \
 -w /work adaptive_cod /bin/bash

Checkpoints with trained model weights

We recommend running python download_weights.py to avoid manually configuring the paths

Official weights for this repo can be found here. They should reside under ./resource/ckpt/acod
Weights from Matsubara et al. can be found here. They should reside under ./resource/ckpt/ghnd
Pretrained weights for Faster/Mask-RCNN are provided by pytorch
Pretrained weights for EfficientDet Models can be found here

Running

Each configuration file at ./config can be used in 3 different ways: train/eval/analyze. If a checkpoint is found at ./resource/ckpt it will be loaded for use.

Obs.: Checkpoint and config file names may be different. They are bound together by config variables.

Examples

Train

python src/mimic_runner.py --config $CONFIG_FILE_PATH -distill

Eval

python src/mimic_runner.py --config $CONFIG_FILE_PATH
python src/eval_dir.py --dir ./config/acod

Analyze

Provides statistics such as operation count, number of parameters and bandwidth requirements

python src/mimic_analyzer.py --config $CONFIG_FILE_PATH
python src/mimic_analyzer.py --dir ./config/acod/

jsiloto / adaptive-cod