SA-AutoAug
Scale-aware Automatic Augmentation for Object Detection
Yukang Chen, Yanwei Li, Tao Kong, Lu Qi, Ruihang Chu, Lei Li, Jiaya Jia
This project provides the implementation for the CVPR 2021 paper "Scale-aware Automatic Augmentation for Object Detection". Scale-aware AutoAug provides a new search space and search metric to find effective data agumentation policies for object detection. It is implemented on maskrcnn-benchmark and FCOS. Both search and training codes have been released. To facilitate more use, we re-implement the training code based on Detectron2.
Installation
For maskrcnn-benchmark code, please follow INSTALL.md for instruction.
For FCOS code, please follow INSTALL.md for instruction.
For Detectron2 code, please follow INSTALL.md for instruction.
Search
(You can skip this step and directly train on our searched policies.)
To search with 8 GPUs, run:
cd /path/to/SA-AutoAug/maskrcnn-benchmark
export NGPUS=8
python3 -m torch.distributed.launch --nproc_per_node=$NGPUS tools/search.py --config-file configs/SA_AutoAug/retinanet_R-50-FPN_search.yaml OURPUT_DIR /path/to/searchlog_dirSince we finetune on an existing baseline model during search, a baseline model is needed. You can download this model for search, or you can use other Retinanet baseline model trained by yourself.
Training
To train the searched policies on maskrcnn-benchmark (FCOS)
cd /path/to/SA-AutoAug/maskrcnn-benchmark
export NGPUS=8
python3 -m torch.distributed.launch --nproc_per_node=$NGPUS tools/train_net.py --config-file configs/SA_AutoAug/CONFIG_FILE OUTPUT_DIR /path/to/traininglog_dirFor example, to train the retinanet ResNet-50 model with our searched data augmentation policies in 6x schedule:
cd /path/to/SA-AutoAug/maskrcnn-benchmark
export NGPUS=8
python3 -m torch.distributed.launch --nproc_per_node=$NGPUS tools/train_net.py --config-file configs/SA_AutoAug/retinanet_R-50-FPN_6x.yaml OUTPUT_DIR models/retinanet_R-50-FPN_6x_SAAutoAugTo train the searched policies on detectron2
cd /path/to/SA-AutoAug/detectron2
python3 ./tools/train_net.py --num-gpus 8 --config-file ./configs/COCO-Detection/SA_AutoAug/CONFIG_FILE OUTPUT_DIR /path/to/traininglog_dirFor example, to train the retinanet ResNet-50 model with our searched data augmentation policies in 6x schedule:
cd /path/to/SA-AutoAug/detectron2
python3 ./tools/train_net.py --num-gpus 8 --config-file ./configs/COCO-Detection/SA_AutoAug/retinanet_R_50_FPN_6x.yaml OUTPUT_DIR output_retinanet_R_50_FPN_6x_SAAutoAugResults
We provide the results on COCO val2017 set with pretrained models.
Based on maskrcnn-benchmark
| Method | Backbone | APbbox | Download |
|---|---|---|---|
| Faster R-CNN | ResNet-50 | 41.8 | Model |
| Faster R-CNN | ResNet-101 | 44.2 | Model |
| RetinaNet | ResNet-50 | 41.4 | Model |
| RetinaNet | ResNet-101 | 42.8 | Model |
| Mask R-CNN | ResNet-50 | 42.8 | Model |
| Mask R-CNN | ResNet-101 | 45.3 | Model |
Based on FCOS
| Method | Backbone | APbbox | Download |
|---|---|---|---|
| FCOS | ResNet-50 | 42.6 | Model |
| FCOS | ResNet-101 | 44.0 | Model |
| ATSS | ResNext-101-32x8d-dcnv2 | 48.5 | Model |
| ATSS | ResNext-101-32x8d-dcnv2 (1200 size) | 49.6 | Model |
Based on Detectron2
| Method | Backbone | APbbox | Download |
|---|---|---|---|
| Faster R-CNN | ResNet-50 | 41.9 | Model - Metrics |
| Faster R-CNN | ResNet-101 | 44.2 | Model - Metrics |
| RetinaNet | ResNet-50 | 40.8 | Model - Metrics |
| RetinaNet | ResNet-101 | 43.1 | Model - Metrics |
| Mask R-CNN | ResNet-50 | 42.9 | Model - Metrics |
| Mask R-CNN | ResNet-101 | 45.6 | Model - Metrics |
Citing SA-AutoAug
Consider cite SA-Autoaug in your publications if it helps your research.
@inproceedings{saautoaug,
title={Scale-aware Automatic Augmentation for Object Detection},
author={Yukang Chen, Yanwei Li, Tao Kong, Lu Qi, Ruihang Chu, Lei Li, Jiaya Jia},
booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2021}
}
Acknowledgments
This training code of this project is built on maskrcnn-benchmark, Detectron2, FCOS, and ATSS. The search code of this project is modified from DetNAS. Some augmentation code and settings follow AutoAug-Det. We thanks a lot for the authors of these projects.
Note that:
(1) We also provides script files for search and training in maskrcnn-benchmark, FCOS, and, detectron2.
(2) Any issues or pull requests on this project are welcome. In addition, if you meet problems when applying the augmentations to other datasets or codebase, feel free to contact Yukang Chen ([email protected]).