Gated Feedback Refinement Network for Dense Image Labeling

Publication

Md Amirul Islam, Mrigank Rochan, Neil D. B. Bruce, Yang Wang
Gated Feedback Refinement Network for Dense Image Labeling, CVPR, 2017.
[BibTeX] [PDF] 
@inproceedings{islamsal18,
  author = {Amirul Islam, Md and Rochan, Mrigank and Bruce, Neil D. B. and Wang, Yang},
  title = {Gated Feedback Refinement Network for Dense Image Labeling},
  booktitle = {Computer Vision and Pattern Recognition (CVPR)},
  year = {2017}
}

Team

Results


Semantic Segmentation:


Quantitative results on the CamVid dataset. We report per-class IoU and mean IoU for each method. Our approach achieves the state-of-the-art results on this dataset. The improvements on smaller and finer objects are particularly pronounced for our model.



Quantitative results in terms of mean IoU on PASCAL VOC 2012 test set. Note that G-FRNet-Res101 includes CRF.



Qualitative results on the CamVid dataset. G-FRNet is capable of retaining the shape of smaller and finer object categories (e.g. column-pole, side-walk, bicyclist, and sign-symbols) accurately.




Qualitative results on PASCAL VOC 2012 validation set




Semantic Object Parsing:

Below are the Comparison of object parsing performance with state-of-the-art methods on Horse-Cow parsing dataset




Coarse-to-Fine Visualization:

Stage-wise visualization of semantic segmentation results on PASCAL VOC 2012. For each row, we show the input image, ground-truth, and the prediction map produced at each stage of our feedback refinement network



Class-wise heatmap visualization on PASCAL VOC 2012 validation set images after each stage of refinement

Downloads