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:

• [FCN-32s sbd]: mean IU 0.601230112927 on aug_val
• [FCN-16s sbd]: mean IU 0.623964674094 on aug_val
• [FCN-8s sbd]: mean IU 0.625525553796 on aug_val
• [FCN-16s OBG-8s sbd]: mean IU 0.628746446579 on aug_val
• [FCN-8s OBG-8s sbd]: mean IU 0.630523623869 on aug_val
• [FCN-8s OBG-4s sbd]: mean IU 0.593030120308 on aug_val
• [FCN-8s OBG-2s sbd]: mean IU 0.577085377376 on aug_val

model link:

• [FCN-8s OBG-2s sbd]: voc-fcn8s-obg2s.caffemodel: https://drive.google.com/open?id=0B5i4atpKg9EcRU9rb1lwd1VnTlE
• [FCN-8s OBG-4s sbd]: voc-fcn8s-obg4s.caffemodel: https://drive.google.com/open?id=0B5i4atpKg9EcU3U5Xy05Tm5kX0U
• [FCN-8s OBG-8s sbd]: voc-fcn8s-obg8s.caffemodel: https://drive.google.com/open?id=0B5i4atpKg9EcMWJXcFl5MGdwQ2c

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