matt-rb / segmentation_cnn_eccv

Image Segmentation Deep-Binary

general image segmentation. Deep-Net to train segments regradless to the object categories.
In preparation for ECCV 2016.

Prerequisites

• Install Torch7
• Have at least Cuda 7.0
• Install CuDNN
• Install matio-ffi

Directories

• model : binary torch alexnet models. to download models run getmodels.sh.

• scripts : includes script and functions for lua, matlab and python.
  lua shared script folder. for import in new script and use functions :

require("scripts.lua.common")
require("scripts.lua.utils")
require("scripts.lua.create_models")

common includes common directories addresses, required libraries and also utils and create_models functions.
utils includes frequently used functions such as load image, converters, etc.
create_models includes functions to create torch models.

matlab matlab script and function folder.

• data : test sample data and images (input data).

Models

to download pretrained models from dropbox run getmodels.sh script from folder /segmentation/model :

$ bash ./getmodels.sh 

getmodels.sh parameters:

$ bash ./getmodels.sh -h
Downloading torch models
-m --model name of model (default all)
-d --dir download directory (default current)
----------------- Model List -------------------------
0 - all
1 - alex_std
2 - alex_fullconv_992
3 - alex_fullconv_1000

Scripts

• th_model_manipulation : First remove the softmax layer, and then fc7 from fully_conv alexnet. creating two new models th_model_full_conv_fc7 and th_model_full_conv_fc6, make them ready to extract fc6, fc7 output feats.

• th_extract_feats : Extract fc6 and fc7 output feats for PASCAL dataset and save to .mat file.