caffe-sal

Caffe-based Amulet and UCF models for salient object detection, shipping training code, pretrained download links, and the citations the authors ask for.

The repository holds the saliency detection models from the authors' ICCV 2017 work. The work looks closely at the convolutional features in pretrained FCN models, and the flagship method, Amulet, aggregates multi-level convolutional features for salient object detection. The README names Amulet and UCF as the two methods provided, with a pair of Google Drive links for the related files, one for each method.

Setup follows Caffe's official site, installing the full toolbox including the necessary Matlab wrappers. The training code lives in the models/Amulet and models/UCF directories, using a batch normalization implementation credited to Alex Kendall and colleagues to save memory. Training follows the Caffe examples, with an ImageData layer loading training image pairs and data lists generated by Matlab, which is a typical workflow for Caffe-era vision repos. The memory-saving BN choice is the notable detail here, since it affects how far you can push batch size on a single GPU.

The saliency maps are downloadable from a link in the results section, and the citation block lists the papers to cite if the models help your research, including the Amulet paper with its authors, Zhang Pingping, Wang Dong, Lu Huchuan, and colleagues, with a second paper credited to Wang Tiantian, Borji Ali, and others. A question and connection section closes the README with a point of contact, which is a common courtesy for research code.

The repo is a practical release of the Amulet and UCF saliency methods, with setup steps, training locations, result downloads, and a citation block all in the README.

• Source 1 · Project repository
• Source 2 · README