jizhihang / weakcnn Goto Github PK

#Version 0.1 This is a matlab implementation of weakly supervised ConvNet, which has its origin from 'Is object localization for free? – Weakly-supervised learning with convolutional neural networks' in CVPR 2015.

##How to implement the code 1.I used matconvnet in this project, which means you should have installed Matconvnet in your computer. Here is the link (http://www.vlfeat.org/matconvnet/) Thanks to A. Vedaldi and K. Lenc for providing such a good tool. You need to download and install Matconvnet.

2.Put the function file in folders 'example' and 'matlab' to corresponding location.(PS: you should have the same two folders in the root path of Matconvnet)

3.run example/cnn_weakly_label.m. ##Some Notes 1.I didn't use GPU here which is easily supported in Matconvnet. So, be free to use it in the code. 2.Global max pooling is implemented in vl_nnglobalmaxpool.m using normal vl_nnpool.m plus some tricks. 3.I added a softmax layer to the original paper. Please read cnn_weakly_label_init.m carefully. 4.Pay attention to the output error. mAP(mean average precision) instead of top error is applied to the output. I didn't modify the output message carefully cause I don't have enough time. This situatio will be improved in the latter version. 5.We only used a small set from PASCAL VOC 2012 which is not enough for complete training process. Feel free to contact me via zhouhy at lamda.nju.edu.cn.

#Version 0.2 This version now supports Gpu acceleration! I modified the loss function and some other details, e.t. random sample. However, I didn't try a multi-scale model. I believe this code could achieve nearly the same accuracy with the original one which is written using torch7.