Introduction

This project is based on libtorch-faster-rcnn, and the motivation is also the same as what's discussed in there. This time, we hope to give one way of how one-stage object detector can be implemented using libtorch.

ATSS

The one-stage object detector we choose here is ATSS. ATSS defines a way to assign positive and negative samples during training, and it can be applied to both anchor-based and anchor-free detectors. Here we implement the algorithm based on anchor-based one-stage detector---RetinaNet. We also add universal tricks like GN, centerness and GIoULoss.

The overall setting and implementation very much follow mmdet's except for the positive-negative sample assignment part, in which this implementation is simpler and easier to follow. However, they produce the same results.

Compile and Use

compile with cmake(v3.19.2)

mkdir build
cd build
cmake .. -DCMAKE_PREFIX_PATH=path_to_libtorch -DCMAKE_PREFIX_PATH=path_to_opencv
cmake --build . --config Release --parallel 8

train

./build/train configs/atss_r50_fpn_voc0712.json --work-dir work_dir --gpu 0

inference

./build/test atss_r50_fpn_voc0712.json epoch_12.pt --out epoch_12.bbox.json --gpu 0

For more guidance please look at readme.

Benchmark

Train: voc2007trainval + voc2012trainval

Test: voc2007test

Results are taking average of three runs.

	backbone	mAP	AP50
mmdet	Resnet50	53.3	78.1
this	Resnet50	53.4	78.1