A simple implements of RCNN with tensorflow

Introduction

This is an experimental Tensorflow implementation of RCNN - a convnet for object detection with a region proposal network. For details about R-CNN please refer to the paper R-CNN: Rich feature hierarchies for accurate object detection and semantic segmentation by Ross Girshick, Jeff Donahue, Trevor Darrell, Jitendra Malik.

Requirements

python 3.6
tensorflow 1.2.1

Content

images/*: Some example imgs to test.
finetune_alexnet/: Storage each epoch's weights in checkpints and log the finetuning progress in TensorBoard.
Models/: To storage alexnet weights and SVM Models.
alexnet.py: an TensorFLow implementation of AlexNet by Alex Krizhevsky at all.
validate_alexnet.py: If you wanna to validate alexnet, this script is helpful(This file has no relation with RCNN). You need are the pretrained weights, which you can find here, place the weights file in 'Models/AlexNet/' and try command "python validat_alexnet.py":
finetune.py、datagenerator.py: Scripts to finetune alexnet, here is the detail instructions to finetune alexnet.
generate_finetune_data.py: To generate the fixed format of train data and val data required by finetune.py. Notice this file only works on VOCdevkit, if you use other data sets, the implementation may not be the same. overall，each of them list the complete path to your train/val images together with the class number in the following structure.

Example train.txt:
/path/to/train/image1.png 0
/path/to/train/image2.png 1
/path/to/train/image3.png 2
/path/to/train/image4.png 0
.
.

ckpt2npy: Change the format form '.ckpy' to '.npy' of alexnet weights.
train_svm.py
tool.py: Some func implements for other files calls, such as Non Maximum Suppression(NMS)、Intersection over Union(IOU)..., and also a func to call module selectivesearch.selective_search, you can try "python tool.py" to check the proposal rectangles generate by 'selective_search':
Note： The proposal rectangles are different according to the argu scale, sigma, min_size in func selective_search().

Usage

Clone the Fast R-CNN repository

git clone https://github.com/asensioatgithub/RCNN.git

Click here to download my tinetuned AlexNet weights and SVM Models, and put them in corresponding directory in "Moldes/".
Install the required python modules, and then instruct the follow cmd:

python demo.py

Train Model

here are my steps to experience RCNN with VOCPascll2012, if you want to train other datasets, they are Referable.

wget the VOCPascll2012 datasets, or click here:

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2011.tar

Extract the tar to repo, so the directory looks like this:

images
Models
VOCtrainval_11-May-2012
alexnet.py
caffe_classes.py
...

Adjust parameter in config.py scenes

class_name = ["bus", "person", "car"] # The classes what rcnn to detect
...

run generate_finetune_data.py to generate finetune_train_list.txt and finetune_val_list.txt, they are looks like:
run finetune.py
run train_svm.py
run demo.py

abao365 / RCNN