Codes for CVPR2016 paper "CRAFT Objects from Images"
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
1_RPN Update README Jun 8, 2016
2_CasRPN fix folder structure under data Jun 8, 2016
3_FRCN Modify 4_CasFRCN Jun 10, 2016
4_CasFRCN Modify 4_CasFRCN Jun 10, 2016
.gitignore add faster_rcnn as 1_RPN Jun 7, 2016 Update results in README Jun 14, 2016


The codes are with the CVPR2016 paper "CRAFT Objects from Images".

In a word, we extend the conventional two-stage object detection framework (first locating object proposals, then classifying object categories) to a four-stage pipeline, in which the proposal localization task is solved with a cascade network of Region Proposal Network (RPN) and Fast R-CNN to improve the proposal quality, while the object classification task is handled by a cascade network of two Fast R-CNN nets with different objective functions (one-hot classification and one-vs-rest classification) to eliminate false positives.

We name our approach "CRAFT" (short for "Cascade Rpn And FasT-rcnn") and show considerable improvement over Fast R-CNN and Faster R-CNN baselines on PASCAL VOC 07/12 and ILSVRC datasets. For more details please refer to our CVPR2016 paper.

The codes are built on RPN (Stage 1) and Fast R-CNN (Stage 2,3,4). It would be easier to use the codes if you are familiar with these two projects.

The codes are tested on Ubuntu 14.04, 256GB Memory, Titan X GPU, MATLAB R2015a.


  1. Follow instructions in Faster R-CNN to make the codes in 1_RPN, using Caffe provided by Shaoqing Ren
  2. Follow instructions in Fast R-CNN to make the codes in 2_CasRPN, 3_FRCN, and 4_CasFRCN, using our slightly modified Caffe
  3. Download the VGG16 pre-trained model and PASCAL VOC 2012 dataset and make proper links pointing to them
  4. You can create a soft link of folders caffe-fast-rcnn and data for 2_CasRPN, 3_FRCN, and 4_CasFRCN for convenience.

Training and testing

The whole pipeline is stage-wise. Now we show how to train an object detector using CRAFT approach on PASCAL VOC 2012 train+val dataset and test it on PASCAL VOC 2012 test set. For simplicity, we do not use joint training between RPN and Fast R-CNN networks.

Stage 1. RPN

cd 1_RPN
matlab ./experiments/script_faster_rcnn_VOC2012_VGG16.m
matlab saveProposals.m

Stage 2. CasRPN

cd 2_CasRPN
matlab saveProposals.m

Stage 3. FRCN

cd 3_FRCN
matlab saveDetections.m

Stage 4. CasFRCN

cd 4_CasFRCN


                      | training data                          | test data            | mAP   

------------------------- |:--------------------------------------:|:--------------------:|:-----: CRAFT, VGG-16 | VOC 2007 trainval + 2012 trainval | VOC 2007 test | 75.7% CRAFT, VGG-16 | VOC 2012 trainval | VOC 2012 test | 71.3%

Note: The real mAP results may vary a little from the above results reported in the paper. We do not adopt joint training between RPN and Fast R-CNN currently.


If you use our codes in your research, we are grateful if you cite the paper:

  title={Craft Objects from Images},
  author={Yang, Bin and Yan, Junjie and Lei, Zhen and Li, Stan},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},


We give our sincere gratitude to the following people, groups and institutions:

  • Anonymous reviewers
  • Ross Girshick for the Fast R-CNN project
  • Shaoqing Ren for the Faster R-CNN project
  • Caffe team
  • VGG team
  • SenseTime Group Limited
  • NVIDIA Corporation