GraphNet
GraphNet: Learning Image Pseudo Annotations for Weakly-Supervised Semantic Segmentation, in ACM Multimedia 2018 (oral) [Paper]
Overview
Implementation Details
The network architecture of DeepLabv2-VGG16 serves as our network architecture of scribble and bounding box annotations experiments.
Results
Results on the PASCAL VOC 2012 dataset
method | annotation | w/o CRF | w/ CRF |
---|---|---|---|
Ours: ScrGraphNet, Initial | scribble | 63.3 | 68.2 |
Ours: ScrGraphNet, 1-Round | scribble | 64.5 | 68.9 |
Ours: BboxGraphNet, Initial | bounding box | 57.1 | 63.4 |
Ours: BboxGraphNet, 1-Round | bounding box | 61.3 | 65.6 |
Results on the PASCAL-CONTEXT dataset
method | w/o CRF | w/ CRF |
---|---|---|
Ours: ScrGraphNet, Initial | 33.1 | 39.7 |
Ours: ScrGraphNet, 1-Round | 33.9 | 40.2 |
Citation
If you find the paper useful for your research, please cite:
@INPROCEEDINGS{Pu2018GraphNet,
author = {Pu, Mengyang and Huang, Yaping and Guan, Qingji and Zou, Qi},
title = {GraphNet: Learning Image Pseudo Annotations for Weakly-Supervised Semantic Segmentation},
booktitle = {ACM Multimedia},
year = {2018}
}