Focal Loss for Dense Rotation Object Detection

Abstract

This repo is based on Focal Loss for Dense Object Detection, and it is completed by YangXue.

This is the baseline work of R³Det, paper link: R³Det: Refined Single-Stage Detector with Feature Refinement for Rotating Object.

Performance

More results and trained models are available in the MODEL_ZOO.md.

DOTA1.0

Model	Backbone	Training data	Val data	mAP	GPU	Image/GPU	Anchor	Reg. Loss	lr schd	Data Augmentation	Configs
RetinaNet (baseline)	ResNet50_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	62.22	1X GeForce RTX 2080 Ti	1	H	smooth L1	1x	No	cfgs_res50_dota_v4.py
RetinaNet (baseline)	ResNet50_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	62.79	8X GeForce RTX 2080 Ti	1	H	smooth L1	2x	No	cfgs_res50_dota_v8.py
RetinaNet (baseline)	ResNet101_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	64.19	1X GeForce RTX 2080 Ti	1	H	smooth L1	1x	No	cfgs_res101_dota_v9.py
RetinaNet (baseline)	ResNet152_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	65.79	8X GeForce RTX 2080 Ti	1	H	smooth L1	2x	No	cfgs_res152_dota_v12.py
RetinaNet (baseline)	ResNet50_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	61.94	1X GeForce RTX 2080 Ti	1	R	smooth L1	1x	No	cfgs_res50_dota_v1.py
RetinaNet (baseline)	ResNet50_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	62.25	8X GeForce RTX 2080 Ti	1	R	smooth L1	2x	No	cfgs_res50_dota_v10.py
RetinaNet	ResNet50_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	68.65	1X GeForce RTX 2080 Ti	1	R	iou-smooth L1	1x	No	cfgs_res50_dota_v5.py
R3Det	ResNet50_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	65.73	8X GeForce RTX 2080 Ti	1	H + R	smooth L1	2x	No	-
R3Det*	ResNet50_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	67.20	8X GeForce RTX 2080 Ti	1	H + R	smooth L1	2x	No	-
R3Det	ResNet101_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	71.69	8X GeForce RTX 2080 Ti	1	H + R	smooth L1	3x	Yes	-
R3Det	ResNet152_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	72.81	8X GeForce RTX 2080 Ti	1	H + R	smooth L1	4x	Yes	-
R3Det*	ResNet152_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	73.74	8X GeForce RTX 2080 Ti	1	H + R	smooth L1	4x	Yes	-
R3Det++	ResNet50_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	68.54	8X GeForce RTX 2080 Ti	1	H + R	smooth L1	2x	No	-
R3Det++	ResNet152_v1d 600->800	DOTA1.0 trainval	DOTA1.0 test	74.41	8X GeForce RTX 2080 Ti	1	H + R	smooth L1	4x	Yes	-
R3Det++	ResNet152_v1d MS	DOTA1.0 trainval	DOTA1.0 test	76.56	4X GeForce RTX 2080 Ti	1	H + R + more	smooth L1	6x	Yes	-

R³Det*: R³Det with two refinement stages
The performance of all models comes from the source paper.

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

IoU-smooth L1 Loss

SCRDet: Towards More Robust Detection for Small, Cluttered and Rotated Objects (ICCV2019)

Download Model

Pretrain weights

1、Please download resnet50_v1, resnet101_v1 pre-trained models on Imagenet, put it to data/pretrained_weights.
2、(Recommend in this repo) 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          

Tensorboard

cd $PATH_ROOT/output/summary
tensorboard --logdir=.

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