This project is trying to Recreate CenternetV2 based on MMDET, which is proposed in paper Probabilistic two-stage detection.

This project is also for the contest OpenMMLab Algorithm Ecological Challenge.

This is NOT the official implementation.

Quick peek at the result:

The paper: centernet2(CascadeRCNN-CenterNet w. prob.) mAP is 42.9.

This implementation: we implement centernet2(CascadeRCNN-CenterNet w. prob.) mAP is 43.2.

Note: We will continue to maintain the code with trick.

Note: We always reproduce the weights, which Verification and inference part.

Code: we add detector in mmdet/models/detectors/centernetv2.py.

Code: we add config in configs/centernetv2/centernet2.py.

Code: we add centernet_head in mmdet/dense_heads/centernet_headv2.py.

Code: we add hm_binary_focal_loss in mmdet/losses/hm_binary_focal_loss.py.

Code: we modify mmdet/roi_heads/cascade_roi_head.py, add the super parameter add_agnostic_score to Control whether the first stage score.This hyperparameter does not affect the use of other configuration files

Code: we modify mmdet/bbox_head/bbox_head.py, add the super parameter add_agnostic_score to Control Whether to use softmax.This hyperparameter does not affect the use of other configuration files

MMDetection: this project is based on version v2.15.0.

MMCV: version v1.3.5

Dataset: coco_train_2017(117k) as training dataset and coco_val_2017(5k) as testing dataset. All the results are reported on coco_val_2017.

Results reported in the paper:

Results by this implementation:

Log and model:

You can train and inference the model like any other models in MMDetection, see docs for details.

conda create -n centernetv2 python=3.7 -y

conda install pytorch cudatoolkit=10.1 torchvision -c pytorch

pip install mmcv-full

git clone https://github.com/yyz561/mmdetection

pip install -r requirements/build.txt

pip install -v -e . # or "python setup.py develop"

Probabilistic two-stage detection

MMDetection

MMCV