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

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

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