silasxue / whats-the-point1

Weakly-supervised semantic image segmentation with CNNs using point supervision

Code for our ECCV paper What's the Point: Semantic Segmentation with Point Supervision.

Summary

This library is a custom build of Caffe for semantic image segmentation with point supervision. It is written for the "FCN-32S-PASCAL" model (fully-convolutional network, stride of 32 for PASCAL VOC 2012), based on this model. More details on the original model are available here.

Quick Start

See the code README page to get started.

Code Structure

All Caffe src files and models are in the caffe directory. All code and scripts to run and evaluate the various models are in the whats-the-point-2016 directory.

Modified Caffe src files:

• softmax_loss_expectation_layer.cpp: CPU version of our custom loss layer
• softmax_loss_expectation_layer.cu: GPU version of our custom loss layer
• loss_layers.hpp: Registers our new loss layer
• caffe.proto: Adds new loss parameters.