yilei0620 / DeepLab-V2-Read-Memory-Data

Modified DeepLab caffe package for reading data from memory

Introduction

DeepLab is a nice framework for image sematic segmentation. Details about DeepLab please refer to the project website built by authors.

However, the original package doesn't provide a method to read data from memory. I found there is a little information in the internet to deal with this issue.

If we want to use MemoryDataLayer to read the data, it doesnt' work in the case that we want to implement crf layer. Thus, I rewrote the MemoryDataLayer to make it output 3 top layers so that now we can use it to read data from memory and feed to Caffe module.

Tutorial

The example can be found in directory example_reading_from_memory. The difference of MemoryDataLayer here is that when we call AddMatVector function, we should have 3 input. The declaration of function is:

void MemoryDataLayer<Dtype>::AddMatVector(const vector<cv::Mat>& mat_vector, const vector<int>& labels, const vector<pair<int, int > >& dim_vector),

where mat_vector is the vector for input images, labels is the vector for corresponding labels and dim_vector is the vector for input images' dimensions. The form of dimensions are <height, width>.