Benzlxs / TFFRCNN

FastER RCNN built on tensorflow

Experiments on how anchor number affect detection results

This repo is forded from https://github.com/CharlesShang/TFFRCNN on implimentation of Faster RCNNFaster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks by Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun.

Code is run to show how anchor number affect detection results,

Acknowledgments:

1. TFFRCNN

Platform

1. Tensorflow 1.7 and Python 2.7
2. Ubuntu 16.04
3. GPU Titan

To repeat the experiments

1. Follow instruction on [TFFRCNN] (https://github.com/CharlesShang/TFFRCNN) to install requirements
2. Download the KITTI dataset and creat dataset folder as instructed in [TFFRCNN] (https://github.com/CharlesShang/TFFRCNN)
3. Download pretrained VGG16 model and locate it in under data folder
4. Design the anchor scale and aspect ratio in the config.py
5. Change the output_dir name in kitti_train.sh and run kitti_train.sh
6. change the name_anchor in all_test.sh and run all_test.sh

Plotting

The plotting result code is located under experiments/plotting folder and written with matlab, after you finish all the experiments, set the correct directory of your result in the plotting_all_data.m. Our experimental results have been uploaded under the data folder