Acknowledgements: a huge thanks to Yangqing Jia for creating Caffe and the BVLC team, and to Ross Girshick for creating RCNN

This is a re-implementation of the object detection algorithm described in the ECCV 2014 paper "Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition". This re-implementation should reproduce the object detection results reported in the paper up to some statistical variance. The models used in the paper are trained/fine-tuned using cuda-convnet, while the model attached with this code is trained/fine-tuned using Caffe, for the ease of code release.

The implementation of image classification training/testing has not been included, but the network configuration files can be found directly in this code.

Please contact [email protected] or [email protected] if you have any question.

If you find SPP_net useful in your research, please consider citing:

@inproceedings{kaiming14ECCV,
    Author = {Kaiming, He and Xiangyu, Zhang and Shaoqing, Ren and Jian Sun},
    Title = {Spatial pyramid pooling in deep convolutional networks for visual recognition},
    Booktitle = {European Conference on Computer Vision},
    Year = {2014}
}

SPP_net is released under the Simplified BSD License for non-commercial use (refer to the LICENSE file for details).

0. Prerequisites
1. MATLAB (tested with 2014a on 64-bit Windows)
2. Caffe's prerequisites (some function is based our modified caffe, so we provied compiled caffe mex and cpp file for mex wapper), run external\fetch_caffe_mex_5_5.m to download
3. News: a caffe version which supports spp mex is provided in https://github.com/ShaoqingRen/caffe/tree/SPP_net, this version is forked from BVLC/caffe on Oct. 1, 2014. For the caffe.mex compiled from this caffe code, the Zeiler CNN network with compatible structure (shared in OneDrive ), and new prototxts for finetune (in ./model-defs) should be used.
4. Install SPP_net
5. Get the SPP_net source code by cloning the repository: git clone https://github.com/ShaoqingRen/SPP_net.git
6. Now change into the SPP_net source code directory
7. SPP_net expects to find Caffe in external/caffe
8. Start MATLAB (make sure you're still in the spp directory): matlab
9. You'll be prompted to download the Selective Search code, which we cannot redistribute. Afterwards, you should see the message SPP_net startup done followed by the MATLAB prompt >>.
10. Run the build script: >> spp_build() (builds liblinear, Selective Search, spp_pool and nms). Don't worry if you see compiler warnings while building liblinear, this is normal on my system.
11. Download the model package by run external\fetch_model_data.m

Let's use PASCAL VOC 2007 as an example. The basic pipeline is:

extract features to disk -> finetune -> train SVMs -> test

You'll need about 20GB of disk space free for the feature cache (which is stored in feat_cache by default. It's best if the feature cache is on a fast, local disk.

An one click script is experiments\Script_spp_voc.m