Receptive Field Block Net for Accurate and Fast Object Detection

By Songtao Liu, Di Huang, Yunhong Wang

Introduction

Inspired by the structure of Receptive Fields (RFs) in human visual systems, we propose a novel RF Block (RFB) module, which takes the relationship between the size and eccentricity of RFs into account, to enhance the discriminability and robustness of features. We further assemble the RFB module to the top of SSD with a lightweight CNN model, constructing the RFB Net detector. You can use the code to train/evaluate the RFB Net for object detection. For more details, please refer to our arXiv paper.

VOC2007 Test

System	mAP	FPS (Titan X Maxwell)
Faster R-CNN (VGG16)	73.2	7
YOLOv2 (Darknet-19)	78.6	40
R-FCN (ResNet-101)	80.5	9
SSD300* (VGG16)	77.2	46
SSD512* (VGG16)	79.8	19
RFBNet300 (VGG16)	80.5	83*
RFBNet512 (VGG16)	82.2	38*

COCO

System	test-dev mAP	Time (Titan X Maxwell)
Faster R-CNN++ (ResNet-101)	34.9	3.36s
YOLOv2 (Darknet-19)	21.6	25ms
SSD300* (VGG16)	25.1	22ms
SSD512* (VGG16)	28.8	53ms
RetinaNet500 (ResNet-101-FPN)	34.4	90ms
RFBNet300 (VGG16)	29.9	15ms*
RFBNet512 (VGG16)	33.8	30ms*
RFBNet512-E (VGG16)	34.4	33ms*

Note: * The speed here is tested on the newest pytorch and cudnn version (0.2.0 and cudnnV6), which is obviously faster than the speed reported in the paper (using pytorch-0.1.12 and cudnnV5).

MobileNet

System	COCO minival mAP	#parameters
SSD MobileNet	19.3	6.8M
RFB MobileNet	20.7*	7.4M

*: slightly better than the original ones in the paper (20.5).

Citing RFB Net

Please cite our paper in your publications if it helps your research:

@article{liu2017RFB,
    title = {Receptive Field Block Net for Accurate and Fast Object Detection},
    author = {Songtao Liu, Di Huang and Yunhong Wang},
    booktitle = {arxiv preprint arXiv:1711.07767},
    year = {2017}
}

Installation
Datasets
Training
Evaluation
Models

Installation

Install PyTorch-0.2.0 by selecting your environment on the website and running the appropriate command.
Clone this repository. This repository is mainly based on ssd.pytorch and Chainer-ssd, a huge thank to them.
- Note: We currently only support Python 3+.
Compile the nms and coco tools:

./make.sh

Note: Check you GPU architecture support in utils/build.py, line 131. Default is:

'nvcc': ['-arch=sm_52',

Install pyinn for MobileNet backbone:

pip install git+https://github.com/szagoruyko/pyinn.git@master

Then download the dataset by following the instructions below and install opencv.

conda install opencv

Note: For training, we currently support VOC and COCO.

Datasets

To make things easy, we provide simple VOC and COCO dataset loader that inherits torch.utils.data.Dataset making it fully compatible with the torchvision.datasets API.

VOC Dataset

Download VOC2007 trainval & test

# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2007.sh # <directory>

Download VOC2012 trainval

# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2012.sh # <directory>

COCO Dataset

Install the MS COCO dataset at /path/to/coco from official website, default is ~/data/COCO. Following the instructions to prepare minival2014 and valminusminival2014 annotations. All label files (.json) should be under the COCO/annotations/ folder. It should have this basic structure

$COCO/
$COCO/cache/
$COCO/annotations/
$COCO/images/
$COCO/images/test2015/
$COCO/images/train2014/
$COCO/images/val2014/

UPDATE: The current COCO dataset has released new train2017 and val2017 sets which are just new splits of the same image sets.

Training

First download the fc-reduced VGG-16 PyTorch base network weights at: https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth or from our BaiduYun Driver
MobileNet pre-trained basenet is ported from MobileNet-Caffe, which achieves slightly better accuracy rates than the original one reported in the paper, weight file is available at: https://drive.google.com/open?id=13aZSApybBDjzfGIdqN1INBlPsddxCK14 or BaiduYun Driver.
By default, we assume you have downloaded the file in the RFBNet/weights dir:

mkdir weights
cd weights
wget https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth

To train RFBNet using the train script simply specify the parameters listed in train_RFB.py as a flag or manually change them.

python train_RFB.py -d VOC -v RFB_vgg -s 300

Note:
- -d: choose datasets, VOC or COCO.
- -v: choose backbone version, RFB_VGG, RFB_E_VGG or RFB_mobile.
- -s: image size, 300 or 512.
- You can pick-up training from a checkpoint by specifying the path as one of the training parameters (again, see train_RFB.py for options)
- If you want to reproduce the results in the paper, the VOC model should be trained about 240 epoches while the COCO version need 130 epoches.

Evaluation

To evaluate a trained network:

python test_RFB.py -d VOC -v RFB_vgg -s 300 --trained_model /path/to/model/weights

By default, it will directly output the mAP results on VOC2007 test or COCO minival2014. For VOC2012 test and COCO test-dev results, you can manually change the datasets in the test_RFB.py file, then save the detection results and submitted to the server.

ArsenLuca / RFBNet