search

jiangruoqiao / RICNN_GongCheng_CVPR2015

Learning Rotation-Invariant Convolutional Neural Networks for Object Detection in VHR Optical Remote Sensing Images
31 stars 9 forks source link

Dataset issue #3

Open nourlaban opened 5 years ago

nourlaban commented 5 years ago

Dear Jiang,

First, I found that the dataset used is not for the paper. the used dataset in the paper is NWPU VHR-10 dataset

Second, I search for the used dataset (rotation-mnist-12k) in the .h5 format but I don't get it.

Third, I try to formulate the dataset using NWPU VHR-10 dataset individual images but it was inconsistent with the rest of the code.

So if you put share a link for the original dataset with the correct format, link for the mnist data set or explain the procedure to how to formulate the correct format for the dataset.

jiangruoqiao commented 5 years ago

OK, I have read your problem. At, first, the H5 format, you must install h5py before you want feed the dataset to RICNN. The how to use the h5py you can Google it. The h5py is a very good format to train CNN, you can make the h5 format by h5py, the original dataset is too big, so I think you should make it by yourself. The method is too simple, you can Google it. In addition, the format of input is (number, width, height, channels, number of rotated).