Focal Loss for Dense Object Detection

Abstract

This is a tensorflow re-implementation of Focal Loss for Dense Object Detection, and it is completed by YangXue.

Performance

Model	Backbone	Training data	Val data	mAP	Inf time (fps)	Model Link	Train Schedule	GPU	Image/GPU	Configuration File
Faster-RCNN	ResNet50_v1 600	VOC07 trainval	VOC07 test	73.09	-	-	-	1X GTX 1080Ti	1	-
FPN	ResNet50_v1 600	VOC07 trainval	VOC07 test	74.26	-	-	-	1X GTX 1080Ti	1	-
RetinaNet	ResNet50_v1 600	VOC07 trainval	VOC07 test	73.16	-	-	-	8X GeForce RTX 2080 Ti	1	cfgs_res50_voc07_v3.py
RetinaNet	ResNet50_v1d 600	VOC07 trainval	VOC07 test	73.26	-	-	-	8X GeForce RTX 2080 Ti	1	cfgs_res50_voc07_v4.py
RetinaNet	ResNet50_v1d 600	VOC07 trainval	VOC07 test	74.00	14.6	model	-	4X GeForce RTX 2080 Ti	2	cfgs_res50_voc07_v5.py
RetinaNet	ResNet50_v1d 600	VOC07+12 trainval	VOC07 test	79.66	-	-	-	8X GeForce RTX 2080 Ti	1	cfgs_res50_voc0712_v1.py
RetinaNet	ResNet101_v1d 600	VOC07+12 trainval	VOC07 test	81.69	-	-	-	8X GeForce RTX 2080 Ti	1	cfgs_res50_voc0712_v4.py
RetinaNet	ResNet101_v1d 800	VOC07+12 trainval	VOC07 test	80.69	-	-	-	8X GeForce RTX 2080 Ti	1	cfgs_res50_voc0712_v3.py
RetinaNet	ResNet50_v1d 600	COCO train2017	COCO val2017 (coco minival)	33.4	12.2	-	1x	8X GeForce RTX 2080 Ti	1	cfgs_res50_coco_1x_v4.py
RetinaNet	ResNet50_v1d 600	COCO train2017	COCO val2017 (coco minival)	34.3 (paper: 34.0)	12.2	model	1x	4X GeForce RTX 2080 Ti	2	cfgs_res50_coco_1x_v5.py

My Development Environment

1、python3.5 (anaconda recommend)
2、cuda9.0
3、opencv(cv2)
4、tfplot (optional)
5、tensorflow >= 1.12

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

Others

1、COCO dataset related

• instances_minival2014.json
• coco_minival2014.odgt
• coco_trainvalmini.odgt

Compile

cd $PATH_ROOT/libs/box_utils/cython_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

cd $PATH_ROOT/data/io/  
python convert_data_to_tfrecord_coco.py --VOC_dir='/PATH/TO/JSON/FILE/' 
                                        --save_name='train' 
                                        --dataset='coco'

3、Multi-gpu train

cd $PATH_ROOT/tools
multi_gpu_train_batch.py

Eval

COCO

cd $PATH_ROOT/tools
python eval_coco.py --eval_data='/PATH/TO/IMAGES/'  
                    --eval_gt='/PATH/TO/TEST/ANNOTATION/'
                    --gpu='0'
                    

cd $PATH_ROOT/tools
python eval_coco_multiprocessing.py --eval_data='/PATH/TO/IMAGES/'  
                                    --eval_gt='/PATH/TO/TEST/ANNOTATION/'
                                    --gpus='0,1,2,3,4,5,6,7'           

PASCAL VOC

cd $PATH_ROOT/tools
python eval.py --eval_dir='/PATH/TO/IMAGES/'  
               --annotation_dir='/PATH/TO/TEST/ANNOTATION/'
               --gpu='0'
                    

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