ATSS-Pytorch

Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection

paper address: https://arxiv.org/pdf/1912.02424.pdf

Requirements

• Linux OS
• Python 3.6 or higher (Python 2 is not supported)
• PyTorch 1.2 or higher
• mmcv 0.4.3
• CUDA 10.0
• GCC(G++) 5.4.0 or higher

Installation

a. Create a conda virtual environment and activate it (Optional but recommended).

conda create --name atss python=3.6
conda activate atss

b. Install pytorch and torchvision.
pip is recommended,

pip install torch==1.2.0 torchvision==0.4.0  #  CUDA 10.0

c. Install mmdet (other dependencies wil be installed automatically).

pip install -r requirements.txt
python setup.py build develop
pip install -v -e .

d. Prepare dataset and checkpoint file.

Download coco dataset and checkpoint file

Fold structure should be as follows:

ATSS-Pytorch
├── data
│   ├── coco
│   │   ├── annotations
│   │   ├── train2017
│   │   ├── val2017
│   │   ├── test2017
├── backbone
│   ├── resnet50-19c8e357.pth

Train

bash scripts/train.sh

Test

bash scripts/test.sh

Results

RetinaNet

[1] 1x and 2x mean the model is trained for 90K and 180K iterations, respectively.
[2] (A=1) mean retinanet only has one 8S square anchor in each level (S is stride)