Home

Awesome

Poly-Scale Convolution

Official implementation of our PSConv operator as described in PSConv: Squeezing Feature Pyramid into One Compact Poly-Scale Convolutional Layer (ECCV'20) by Duo Li, Anbang Yao and Qifeng Chen on the MS COCO 2017 benchmark.

<p align="center"><img src="fig/psconv.png" width="600" /></p>

We collect multi-scale feature representations in a finer granularity, by tactfully allocating a spectrum of dilation rates in the kernel lattice.

Getting Started

Installation

Following the instructions from INSTALL.md for installation. More detailed guidance can be found from MMDetection.

Download ImageNet pre-trained checkpoints

Fetch pre-trained weights of PS-ResNet-50, PS-ResNet-101 and PS-ResNeXt-101 (32x4d) backbones and put them to your local path. Set the pretrained path of config file and launch the training of detectors.

Training

The default learning rate in config files is for 8 GPUs and 2 img/GPU (batch size = 8*2 = 16). According to the Linear Scaling Rule, you need to set the learning rate proportional to the batch size if you use different GPUs or images per GPU, e.g., lr=0.01 for 4 GPUs * 2 img/gpu and lr=0.08 for 16 GPUs * 4 img/GPU.

# single-gpu training
python tools/train.py ${CONFIG_FILE}

# multi-gpu training
./tools/dist_train.sh ${CONFIG_FILE} ${GPU_NUM} [optional arguments]

Optional arguments are:

Test

# single-gpu testing
python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [--out ${RESULT_FILE}] [--eval ${EVAL_METRICS}] [--show]

# multi-gpu testing
./tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM} [--out ${RESULT_FILE}] [--eval ${EVAL_METRICS}]

Optional arguments:

Model Zoo

Faster R-CNN

BackboneStyleLr schdbox APDownload
R-50-FPNpytorch1x38.4model | log
R-101-FPNpytorch1x40.9model | log
X-101-FPNpytorch1x41.3model | log

Mask R-CNN

BackboneStyleLr schdbox APmask APDownload
R-50-FPNpytorch1x39.435.6model | log
R-101-FPNpytorch1x41.637.4model | log
X-101-FPNpytorch1x42.438.0model | log

Cascade R-CNN

BackboneStyleLr schdbox APDownload
R-50-FPNpytorch1x41.9model | log
R-101-FPNpytorch1x43.8model | log
X-101-FPNpytorch1x44.4model | log

Cascade Mask R-CNN

BackboneStyleLr schdbox APmask APDownload
R-50-FPNpytorch1x42.936.9model | log
R-101-FPNpytorch1x44.638.4model | log
X-101-FPNpytorch1x45.338.9model | log

Acknowledgement

This implementation is built upon MMDetection. Thanks Kai Chen for releasing this awesome toolbox and his helpful discussions.

Since this project is finished nearly one year ago, we adapt our code from an early commit 713e98b.

Citation

If you find our work useful in your research, please consider citing:

@InProceedings{Li_2020_ECCV,
author = {Li, Duo and Yao, Anbang and Chen, Qifeng},
title = {PSConv: Squeezing Feature Pyramid into One Compact Poly-Scale Convolutional Layer},
booktitle = {The European Conference on Computer Vision (ECCV)},
month = {August},
year = {2020}
}