Faster R-CNN in PyTorch

中文

This is a simple implementation of Faster R-CNN. I mainly referred to the following repositories:

• py-faster-rcnn
• faster-rcnn.pytorch

Progress

• Training and testing on VOC
• VGG16
• Training and testing on COCO
• ResNet

Prerequisites

• python 3
• pytorch 1.0
• python-opencv
• matplotlib
• yacs

Installation

1. Clone the repository

git clone https://github.com/DHZS/faster-rcnn-pytorch.git

2. Compile modules

cd faster-rcnn-pytorch/model/cpp
python setup.py build develop

Training on PASCAL VOC

Prepare the data

Configure your dataset path in yaml config file. e.g.

# Train on VOC07&12
dataset_name: ['voc_2007', 'voc_2012']
dataset_type: ['voc', 'voc']
dataset_path: ['/dataset/pascal_voc/VOC2007', '/dataset/pascal_voc/VOC2012']
image_set: ['trainval', 'trainval']
use_difficult: [True, True]

Download the model pretrained on ImageNet

VGG16_no_bn(Google Drive)

Configure the model file path. e.g.

train:
  base_net: './output/faster_rcnn_base_vgg16_caffe_no_bn.pth'

Train & Evaluation & Visualization

# train
python do_train.py --config_path=config/_train_voc.yaml

# evaluation
python do_eval.py --config_path=config/_eval_voc.yaml

# visualization
python do_visual.py --config_path=config/_visual_voc.yaml

Result

VGG16(conv5_3), Roi Align, VOC07 trainval/VOC07 test

V0C07 metric? Yes
AP for aeroplane = 0.7222
AP for bicycle = 0.7745
AP for bird = 0.6855
AP for boat = 0.5490
AP for bottle = 0.5627
AP for bus = 0.7931
AP for car = 0.8554
AP for cat = 0.8283
AP for chair = 0.4850
AP for cow = 0.7905
AP for diningtable = 0.6418
AP for dog = 0.7935
AP for horse = 0.8343
AP for motorbike = 0.7754
AP for person = 0.7725
AP for pottedplant = 0.4553
AP for sheep = 0.6845
AP for sofa = 0.6376
AP for train = 0.7485
AP for tvmonitor = 0.7229
Mean AP = 0.7056