Faster RCNN

• Author: Haoxin Lin
• Data: 03/05/2020
• Version: 1.0.0
• Brief: A Faster RCNN complemented with TensorFlow.

Test Environment

• OS Environ
  Ubuntu 16.04.5
GPU: TITAN X (Pascal)
Anaconda 4.8.2
CUDA 10.2
• Dependencies
  python 3.6.10
tensorflow-gpu 1.15.0
numpy 1.18.1
Pillow 5.3.0

Usage

1. Download VOC2007 dataset with following instructions.

   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar

2. Extract all of these tars into one directory named VOCdevkit.

   tar xvf VOCtrainval_06-Nov-2007.tar
   tar xvf VOCtest_06-Nov-2007.tar
   tar xvf VOCdevkit_08-Jun-2007.tar

3. Set data at the correct path of Faster RCNN project.

   mv $VOCdevkit $Faster-RCNN/data/VOCdevkit2007

4. Download pre-trained ImageNet VGG16 model.

   cd $Faster-RCNN/model/pretrained
   wget http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz
   tar xvf vgg_16_2016_08_28.tar.gz
   rm vgg_16_2016_08_28.tar.gz

5. It should have this basic structure.

   Faster RCNN
       |──data
       |    |──VOCdevkit2007
       |    |    |──VOCcode
       |    |    |──VOC2007
       |    |    └──···
       |    └──···
       |──model
       |    └──pretrained
       |        └──vgg_16.ckpt
       |──network
       |    └──···
       |──train.py
       └──test.py
   

6. Train Faster RCNN with Pascal VOC 2007.

   python3 train.py

7. Test.

   python3 test.py

Results

class	AP	class	AP
aeroplane	0.6974	diningtable	0.6314
bicycle	0.7100	dog	0.7518
bird	0.5914	horse	0.8008
boat	0.5430	motorbike	0.6987
bottle	0.4067	person	0.7025
bus	0.7478	pottedplant	0.3567
car	0.7160	sheep	0.5993
cat	0.7687	sofa	0.6431
chair	0.4441	train	0.6885
cow	0.7383	tvmonitor	0.5875

Mean AP = 0.6412