Anchor pruning for object detection [CVIU 2022] [arXiv]

By Maxim Bonnaerens, Matthias Freiberger and Joni Dambre.

Abstract

This paper proposes anchor pruning for object detection in one-stage anchor-based detectors. While pruning techniques are widely used to reduce the computational cost of convolutional neural networks, they tend to focus on optimizing the backbone networks where often most computations are. In this work we demonstrate an additional pruning technique, specifically for object detection: anchor pruning. With more efficient backbone networks and a growing trend of deploying object detectors on embedded systems where post-processing steps such as non-maximum suppression can be a bottleneck, the impact of the anchors used in the detection head is becoming increasingly more important. In this work, we show that many anchors in the object detection head can be removed without any loss in accuracy. With additional retraining, anchor pruning can even lead to improved accuracy. Extensive experiments on SSD and MS COCO show that the detection head can be made up to 44% more efficient while simultaneously increasing accuracy. Further experiments on RetinaNet and PASCAL VOC show the general effectiveness of our approach. We also introduce ` overanchorized' models that can be used together with anchor pruning to eliminate hyperparameters related to the initial shape of anchors.

Citation

@article{bonnaerens2022anchor,
  title={Anchor pruning for object detection},
  author={M. Bonnaerens, M. Freiberger and J. Dambre},
  journal={Computer Vision and Image Understanding},
  pages={103445},
  year={2022},
  publisher={Elsevier},
  doi = {https://doi.org/10.1016/j.cviu.2022.103445},
}

Results and models of SSD

Anchor Configuration	AP .50:.95	FLOPS head	BBoxes	Config	Download
SSD Baseline	25.6	4231M	8732	config	model
SSD Configuration-A retrained	25.4	3607M	7814	config	model
SSD Configuration-B retrained	25.6	2476M	4926	config	model
SSD Configuration-C retrained	25.2	1628M	3121	config	model
SSDConfiguration-D retrained	22.8	774M	1291	config	model
RetinaNet Baseline	36.5	129B		config	model
RetinaNet Pruned	34.8	31B		config	model

Above results are on the COCO validation set while the results in the paper are on the COCO test set.

Installation

This repository builds upon MMDetection.

See The MMDetection documentation for installation instructions. Last confirmed working version is mmdet v2.25.0 with mmcv-full v1.4.8

pip install openmim
mim install mmcv-full==1.4.8
git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
git switch --detach v2.25.0
pip install -v -e .

Next, clone our repository and install the anchor pruning package

git clone https://github.com/Mxbonn/anchor_pruning.git
cd anchor_pruning
pip install -e .

Getting started.

Please see Tutorial.ipynb for a general guide on how to do anchor pruning.

To run the given pretrained models above run

python tools/mmdet_test.py configs/ssd/configuration_B.py pretrained_models/configuration_B.pth --eval bbox

after modifying the paths to the mmdet base config in configuration_X.py and linking your dataset directory to data/ similarly as required for mmdetection.