A Pytorch Implementation of R-FCN/CoupleNet

Introduction

This project is an pytorch implement R-FCN and CoupleNet, large part code is reference from jwyang/faster-rcnn.pytorch. The R-FCN structure is refer to Caffe R-FCN and Py-R-FCN

For R-FCN, mAP@0.5 reached 73.2 in VOC2007 trainval dataset
For CoupleNet, mAP@0.5 reached 75.2 in VOC2007 trainval dataset

R-FCN

arXiv:1605.06409: R-FCN: Object Detection via Region-based Fully Convolutional Networks

This repo has following modification compare to jwyang/faster-rcnn.pytorch:

R-FCN architecture: We refered to the origin [Caffe version] of R-FCN, the main structure of R-FCN is show in following figure.
PS-RoIPooling with CUDA :(refer to the other pytorch implement R-FCN, pytorch_RFCN). I have modified it to fit multi-image training (not only batch-size=1 is supported)
Implement multi-scale training: As the original paper says, each image is randomly reized to differenct resolutions (400, 500, 600, 700, 800) when training, and during test time, we use fix input size(600). These make 1.2 mAP gain in our experiments.
Implement OHEM: in this repo, we implement Online Hard Example Mining(OHEM) method in the paper, set OHEM: False in cfgs/res101.yml for using OHEM. Unluckly, it cause a bit performance degration in my experiments

CoupleNet

arXiv:1708.02863:CoupleNet: Coupling Global Structure with Local Parts for Object Detection

Making changes based on R-FCN
Implement local/global FCN in CoupleNet

Tutorial

R-FCN blog

Benchmarking

We benchmark our code thoroughly on three datasets: pascal voc using two different architecture: R-FCN and CoupleNet. Results shows following:

1). PASCAL VOC 2007 (Train: 07_trainval - Test: 07_test, scale=400, 500, 600, 700, 800)

model	#GPUs	batch size	lr	lr_decay	max_epoch	time/epoch	mem/GPU	mAP
R-FCN	1	2	4e-3	8	7	0.88 hr	3000 MB	73.2
CouleNet	1	2	4e-3	8	10	0.60 hr	8900 MB	75.2

Preparation

First of all, clone the code

$ git clone https://github.com/princewang1994/R-FCN.pytorch.git

Then, create a folder:

$ cd R-FCN.pytorch && mkdir data
$ ln -s $VOC_DEVKIT_ROOT .

prerequisites

Python 3.6
Pytorch 0.3.0 or higher
CUDA 8.0 or higher

Data Preparation

PASCAL_VOC 07+12: Please follow the instructions in py-faster-rcnn to prepare VOC datasets. Actually, you can refer to any others. After downloading the data, creat softlinks in the folder data/.

Compilation

As pointed out by ruotianluo/pytorch-faster-rcnn, choose the right -arch in make.sh file, to compile the cuda code:

GPU model	Architecture
TitanX (Maxwell/Pascal)	sm_52
GTX 960M	sm_50
GTX 1080 (Ti)	sm_61
Grid K520 (AWS g2.2xlarge)	sm_30
Tesla K80 (AWS p2.xlarge)	sm_37

More details about setting the architecture can be found here or here

Install all the python dependencies using pip:

$ pip install -r requirements.txt

Compile the cuda dependencies using following simple commands:

$ cd lib
$ sh make.sh

It will compile all the modules you need, including NMS, ROI_Pooing, ROI_Align and ROI_Crop. The default version is compiled with Python 2.7, please compile by yourself if you are using a different python version.

Train

To train a faster R-CNN model with vgg16 on pascal_voc, simply run:

$ CUDA_VISIBLE_DEVICES=$GPU_ID python trainval_net.py \
				   --arch rfcn \
                   --dataset pascal_voc --net res101 \
                   --bs $BATCH_SIZE --nw $WORKER_NUMBER \
                   --lr $LEARNING_RATE --lr_decay_step $DECAY_STEP \
                   --cuda

Set --s to identified differenct experiments.
For CoupleNet training, replace --arch rfcn with --arch couplenet, other arguments should be modified according to your machine. (e.g. larger learning rate for big batch-size)
Model are saved to $RFCN_ROOT/save

Test

If you want to evlauate the detection performance of a pre-trained model on pascal_voc test set, simply run

$ python test_net.py --dataset pascal_voc --arch rfcn \
				   --net res101 \
                   --checksession $SESSION \
                   --checkepoch $EPOCH \
                   --checkpoint $CHECKPOINT \
                   --cuda

Specify the specific model session(--s in training phase), chechepoch and checkpoint, e.g., SESSION=1, EPOCH=6, CHECKPOINT=416.

Demo

Below are some detection results:

Going to do

Keeping update structures to reach the state-of-art
More benchmarking in VOC0712/COCO
Pretrained model for VOC

Acknowledgement

This project is equally writen by Prince Wang, but thanks the faster-rcnn.pytorch's code provider jwyang

zorrocai/R-FCN.pytorch