NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Detection

Abstract

This repo is based on FPN, and completed by YangXue.

Train on COCO train2017 and test on COCO val2017 (coco minival).

COCO

Model	Backbone	Train Schedule	GPU	Image/GPU	FP16	Box AP
Faster (Face++ & Detectron)	R50v1-FPN	1X	8X TITAN Xp	2	no	36.4
Faster (SimpleDet)	R50v1-FPN	1X	8X 1080Ti	2	no	36.5
Faster (ours)	R50v1-FPN	1X	1X TITAN Xp	1	no	36.1
Faster (ours)	R50v1-FPN	1X	4X TITAN Xp	1	no	36.1

Model	Backbone	Pyramid method	Train Schedule	GPU	Image/GPU	Stacks	Dimension	3x3 relu	Box AP
Faster (ours)	R50v1	FPN	1X	4X TITAN Xp	1	0	256	no	36.1
Faster (ours)	R50v1	FPN	1X	8X 2080Ti	1	3	256	yes	35.8
Faster (ours)	R50v1	NAS-FPN	1X	8X 2080Ti	1	3	256	yes	37.9
Faster (ours)	R50v1	NAS-FPN	1X	8X 2080Ti	1	7	256	yes	38.1
Faster (ours)	R50v1	NAS-FPN	1X	8X 2080Ti	1	7	384	yes	38.9

My Development Environment

1、python3.5 (anaconda recommend)
2、cuda9.0 (If you want to use cuda8, please set CUDA9 = False in the cfgs.py file.)
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) Or you can choose to use a better backbone, refer to gluon2TF.

• Baidu Drive, password: 5ht9.
• Google Drive

Trained weights

Select a configuration file in the folder ($PATH_ROOT/libs/configs/) and copy its contents into cfgs.py, then download the corresponding weights.

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
python multi_gpu_train.py

Eval

cd $PATH_ROOT/tools
python eval_coco.py --eval_data='/PATH/TO/IMAGES/'  
                    --eval_gt='/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/CharlesShang/FastMaskRCNN
5、https://github.com/matterport/Mask_RCNN
6、https://github.com/msracver/Deformable-ConvNets