Arbitrary-Oriented Object Detection with Circular Smooth Label
Abstract
This repo is based on Focal Loss for Dense Object Detection, and it is completed by YangXue.
Pipeline
Circular Smooth Label
Performance
More results and trained models are available in the MODEL_ZOO.md.
DOTA1.0 (Task1)
Model | Backbone | Training data | Val data | mAP | Model Link | Anchor | Label Mode | Reg. Loss | Angle Range | lr schd | Data Augmentation | GPU | Image/GPU | Configs |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
CSL | ResNet50_v1 600->800 | DOTA1.0 trainval | DOTA1.0 test | 39.52 | - | H | Pulse | smooth L1 | 90 | 1x | × | 1X GeForce RTX 2080 Ti | 1 | cfgs_res50_dota_v20.py |
CSL | ResNet50_v1 600->800 | DOTA1.0 trainval | DOTA1.0 test | 58.86 | - | H | Rectangular | smooth L1 | 90 | 1x | × | 1X GeForce RTX 2080 Ti | 1 | cfgs_res50_dota_v21.py |
CSL | ResNet50_v1 600->800 | DOTA1.0 trainval | DOTA1.0 test | 60.15 | - | H | Triangle | smooth L1 | 90 | 1x | × | 1X GeForce RTX 2080 Ti | 1 | cfgs_res50_dota_v22.py |
CSL | ResNet50_v1 600->800 | DOTA1.0 trainval | DOTA1.0 test | 63.51 | - | H | Gaussian | smooth L1 | 90 | 2x | × | 2X GeForce RTX 2080 Ti | 1 | cfgs_res50_dota_v18.py |
CSL | ResNet50_v1 600->800 | DOTA1.0 trainval | DOTA1.0 test | 42.06 | - | H | Pulse | smooth L1 | 180 | 2x | × | 4X GeForce RTX 2080 Ti | 1 | cfgs_res50_dota_v28.py |
CSL | ResNet50_v1 600->800 | DOTA1.0 trainval | DOTA1.0 test | 61.98 | - | H | Rectangular | smooth L1 | 180 | 2x | × | 2X GeForce RTX 2080 Ti | 1 | cfgs_res50_dota_v23.py |
CSL | ResNet50_v1 600->800 | DOTA1.0 trainval | DOTA1.0 test | 57.94 | - | H | Triangle | smooth L1 | 180 | 2x | × | 4X GeForce RTX 2080 Ti | 1 | cfgs_res50_dota_v26.py |
CSL | ResNet50_v1 600->800 | DOTA1.0 trainval | DOTA1.0 test | 64.50 | - | H | Gaussian | smooth L1 | 180 | 2x | × | 2X Quadro RTX 8000 | 1 | cfgs_res50_dota_v27.py |
CSL | ResNet50_v1 600->800 | DOTA1.0 trainval | DOTA1.0 test | 65.09 | - | H | Gaussian | smooth L1 + atan(theta) | 180 | 2x | × | 2X Quadro RTX 8000 | 1 | cfgs_res50_dota_v31.py |
CSL | ResNet152_v1 MS | DOTA1.0 trainval | DOTA1.0 test | 70.29 | model | H | Gaussian | smooth L1 + atan(theta) | 180 | 2x | √ | 2X Quadro RTX 8000 | 1 | cfgs_res152_dota_v36.py |
Visualization
My Development Environment
docker images: docker pull yangxue2docker/yx-tf-det:tensorflow1.13.1-cuda10-gpu-py3
1、python3.5 (anaconda recommend)
2、cuda 10.0
3、opencv(cv2)
4、tfplot 0.2.0 (optional)
5、tensorflow 1.13
Download Model
Pretrain weights
1、Please download resnet50_v1, resnet101_v1 pre-trained models on Imagenet, put it to data/pretrained_weights.
2、(Recommend) Or you can choose to use a better backbone, refer to gluon2TF.
- Baidu Drive, password: 5ht9.
- Google Drive
Compile
cd $PATH_ROOT/libs/box_utils/cython_utils
python setup.py build_ext --inplace (or make)
cd $PATH_ROOT/libs/box_utils/
python setup.py build_ext --inplace
Train
1、If you want to train your own data, please note:
(1) Modify parameters (such as CLASS_NUM, DATASET_NAME, VERSION, etc.) in $PATH_ROOT/libs/configs/cfgs.py
(2) Add category information in $PATH_ROOT/libs/label_name_dict/lable_dict.py
(3) Add data_name to $PATH_ROOT/data/io/read_tfrecord.py
2、Make tfrecord
For DOTA dataset:
cd $PATH_ROOT\data\io\DOTA
python data_crop.py
cd $PATH_ROOT/data/io/
python convert_data_to_tfrecord.py --VOC_dir='/PATH/TO/DOTA/'
--xml_dir='labeltxt'
--image_dir='images'
--save_name='train'
--img_format='.png'
--dataset='DOTA'
3、Multi-gpu train
cd $PATH_ROOT/tools
python multi_gpu_train.py
Eval
cd $PATH_ROOT/tools
python test_dota.py --test_dir='/PATH/TO/IMAGES/'
--gpus=0,1,2,3,4,5,6,7
--s (visualization, optional)
--ms (multi-scale test, optional)
Tensorboard
cd $PATH_ROOT/output/summary
tensorboard --logdir=.
Citation
If this is useful for your research, please consider cite.
@article{yang2020arbitrary,
title={Arbitrary-Oriented Object Detection with Circular Smooth Label},
author={Yang, Xue and Yan, Junchi},
journal={arXiv preprint arXiv:2003.05597},
year={2020}
}
@inproceedings{xia2018dota,
title={DOTA: A large-scale dataset for object detection in aerial images},
author={Xia, Gui-Song and Bai, Xiang and Ding, Jian and Zhu, Zhen and Belongie, Serge and Luo, Jiebo and Datcu, Mihai and Pelillo, Marcello and Zhang, Liangpei},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
pages={3974--3983},
year={2018}
}
Reference
1、https://github.com/endernewton/tf-faster-rcnn
2、https://github.com/zengarden/light_head_rcnn
3、https://github.com/tensorflow/models/tree/master/research/object_detection
4、https://github.com/fizyr/keras-retinanet