Codes for Weakly Supervised Saliency Detection

Contact: Kuang-Jui Hsu

Last update: 2017/12/30

Platform: Ubuntu 14.04, MatConvnet 1.0-beta24 (We don's support any installation problem of MatConvnet.)

---

Paper: [BMVC17] Weakly Supervised Saliency Detection with A Category-Driven Map Generator

Authors: Kuang-Jui Hsu, Yen-Yu Lin, Yung-Yu Chuang

PDF: Link1, Link2

Please cite our paper if this code is useful for your research.

@inproceedings{HsuBMVC17,
  author = {K.-J. Hsu and Y.-Y. Lin and Y.-Y Chuang},
  booktitle = {British Machine Vision Conference (BMVC)},
  title = {Weakly Supervised Saliency Detection with A Category-Driven Map Generator},
  year = {2017}
}

---

Demo for training and test: "Run.m"

• This code is only for demo and different from the origianl code because some files were overwrited when I worked for the journal extension.

• The output size of the generator is W * H * 1 after a sigmoid normalization instead of W * H * 2 after a softmax normalization.

• The parameters for weighting losses are not tuned.

• The random seed and the number of total epoches may effect the performance.

• The results are slightly better than ones reported in [BMVC'17] for Graz02 Dataset.

  • Link for Epoches
  • Link for Results
  • Graz02 Results with the release code

Bike	Car	Person	Mean
80.4	63.1	66.5	70.0

---

Results used in BMVC'17:

• Graz02
• VOC07
• VOC12

---

Link: Some results in the journal extension (will be released until it is accepted).

• Faster speed:

• Better results: