dennisleouts/aprcnn_cvpr16

forked from Zhe Xu from UTS

Weakly-supervised Dataset augmented Part-based RCNNs for Fine-grained Category Detection

This work is created by Zhe Xu from UTS. Credit to Ning Zhang, Jeff Donahue, Ross Girshick and Trevor Darrell from UC Berkeley who share the original RCNN and Part-based RCNN code.

Prerequisites

0. Caffe

• Download caffe from http://caffe.berkeleyvision.org/ and follow the instructions to install.
• Change caffe matlab wrapper path in init.m

0. RCNN

• A modified RCNN code is attached in the code patch.
• If you want to access the original RCNN code to see the modifications, download source code from https://github.com/rbgirshick/rcnn and follow the instructions to install.
• Follow rcnn instructions to train the part detectors.

0. Liblinear

• Download liblinear package from http://www.csie.ntu.edu.tw/~cjlin/liblinear/
• Download MALSAR package from http://www.public.asu.edu/~jye02/Software/MALSAR/

Annotation/ has annotated part boxes on CUB200-2011 dataset.

Cached Files

• Recommended file: https://drive.google.com/file/d/0Bwo0SFiZwl3JR1RFdnM0R3EyNDA/view
• Full file: https://drive.google.com/file/d/0Bwo0SFiZwl3JSmZNXzZJVG5FbEE/view

Usage

• If you have installed cached files, run run_prcnn_sample.m to get results
• To run the whole procedure of extracting selective search data -> finetune part cnns -> train RCNN part detectors -> get augmented training data from weak dataset -> re-finetune part cnns -> get final classification results, go to ./rcnn directory and run rcnn_run.m.
• Modify the paths in config_prcnn.m and get_bird_data.m

###Bug report
If you have any issues running the codes, please contact Zhe Xu ([email protected]).