Pytorch Faster-RCNN

Introduction

This project is aimed to reproduce the faster rcnn object detection model. It is developed based on the following projects:

1. rbgirshick/py-faster-rcnn, developed based on Pycaffe + Numpy

2. longcw/faster_rcnn_pytorch, developed based on Pytorch + Numpy

3. endernewton/tf-faster-rcnn, developed based on TensorFlow + Numpy

4. ruotianluo/pytorch-faster-rcnn, developed based on Pytorch + TensorFlow + Numpy

However, there are several unique features compared with the above implementations:

1. It is pure Pytorch code. We convert all the numpy implementations to pytorch.

2. It supports trainig batchsize > 1. We revise all the layers, including dataloader, rpn, roi-pooling, etc., to train with multiple images at each iteration.

3. It supports multiple GPUs. We use a multiple GPU wrapper (nn.DataParallel here) to make it flexible to use one or more GPUs, as a merit of the above two features.

4. It is memory efficient. We limit the image aspect ratio, and group the image in batch with similar aspect ratio. We can train resnet101 and VGG16 with batchsize = 4 (4 images) on a sigle Titan X 12 GB. When training with 8 GPU, the maximum batchsize for each GPU is 3 images (Res101), with total batchsize = 24.

5. It is faster. With above merits, our training speed can achieve xxx / xxx (VGG/Res101) on single TITAN X Pascal GPU and xxx/xxx (VGG / Res101) on 8 TITAN X Pascal GPU.

Benchmarking

We benchmark our code thoroughly on three datasets: pascal voc, mscoco and imagenet-200, using two different network architecture: vgg16 and resnet101. Below are the results:

1. PASCAL VOC

   model	Train Set	Test Set	GPUs	Batch Size	Speed	Memory	mAP
   VGG-16	07trainval	07test	1 Titan X	1	0	0	0
   VGG-16	07trainval	07test	1 Titan X	4	0	0	0
   VGG-16	07trainval	07test	8 Titan X	27	0	0	0
   Res-101	07trainval	07test	1 Titan X	1	0	0	0
   Res-101	07trainval	07test	1 Titan X	4	0	0	0
   Res-101	07trainval	07test	8 Titan X	27	0	0	0
   -----------	-----------	-----------	----------	------------	-------	--------	-----
   VGG-16	coco_train	coco_test	1 Titan X	1	0	0	0
   VGG-16	coco_train	coco_test	1 Titan X	4	0	0	0
   VGG-16	coco_train	coco_test	8 Titan X	27	0	0	0
   Res-101	coco_train	coco_test	1 Titan X	1	0	3085	0
   Res-101	coco_train	coco_test	1 Titan X	4	0	0	0
   Res-101	coco_train	coco_test	8 Titan X	27	0	0	0

Prepare Data

PASCAL_VOC and COCO:

Please follow the instructions of py-faster-rcnn to setup VOC and COCO datasets. The steps involve downloading data and optionally creating softlinks in the data folder. Since faster RCNN does not rely on pre-computed proposals, it is safe to ignore the steps that setup proposals.

ImageNet:

To train a resnet101, run:

 CUDA_VISIBLE_DEVICES=0 python trainval_net.py --dataset pascal_voc --net resnet101

Alternatively, to train a vgg16, run:

CUDA_VISIBLE_DEVICES=0 python trainval_net.py --dataset pascal_voc --net vgg16