Image Segmentation Deep-Binary
general image segmentation. Deep-Net to train segments regradless to the object categories.
In preparation for ECCV 2016.
Prerequisites
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
andth_model_full_conv_fc6
, make them ready to extract fc6, fc7 output feats. -
th_extract_feats : Extract
fc6
andfc7
output feats for PASCAL dataset and save to.mat
file.