mmdet-centernetv2
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.
Implementation
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
Experiments
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:
| AP | AP50 | AP75 | APs | APm | APl | |
|---|---|---|---|---|---|---|
| cascadeRCNN-CenterNet w.prob | 42.862 | 59.519 | 47.028 | 24.064 | 47.043 | 56.197 |
Results by this implementation:
| AP | AP50 | AP75 | APs | APm | APl | |
|---|---|---|---|---|---|---|
| cascadeRCNN-CenterNet w.prob | 43.2 | 60.6 | 47.9 | 25.3 | 46.6 | 56.2 |
Log and model:
| backbone | Iter | bbox AP | Config | Log | Model | GPUs | |
|---|---|---|---|---|---|---|---|
| cascadeRCNN-CenterNet w.prob | R-50-FPN | 90000 | 43.2 | config | log | baidu [jip5] | single-v100(batch=20) |
Usage
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"
Acknowledgement
Probabilistic two-stage detection
