twtygqyy / obg_fcn

3 stars 1 forks source link

OBG-FCN

This repository is to reproduce the implementation of 'Object Boundary Guided Semantic Segmentation' in http://arxiv.org/abs/1603.09742

Object Boundary Guided Semantic Segmentation
Qin Huang, Chunyang Xia, Wenchao Zheng, Yuhang Song, Hao Xu, C.-C. Jay Kuo
arXiv:1603.09742

the paper claimed to achieve 87.5% mean IU in PASCAL VOC 2011 validation set with only the training images of VOC 2011 training set.

The code is based on the repository of https://github.com/shelhamer/fcn.berkeleyvision.org, which contains the offical code for the paper:

Fully Convolutional Models for Semantic Segmentation
Jonathan Long*, Evan Shelhamer*, Trevor Darrell
CVPR 2015
arXiv:1411.4038

The implementation is just for test and could not achieve result close to Object Boundary Guided Semantic Segmentation so far. Any suggestion is more than welcome

Mdoels are trained using extra data from Hariharan et al., but excluding SBD val. Mdoels are tested using aug_val set by excluding the overlapping images in VOC train_val dataset.

Here is the result so far:

model link:

There must be major bugs in the implementation since the performace decreased when combining pool2 and pool1 for object boundary.