SalFBNet Pseudo Saliency Dataset
We create a large-scale Pseudo-Saliency dataset (SalFBNet Pseudo Saliency Dataset) to alleviate the problem of data deficiency in saliency detection. We postulate that existing state-of-the-art pre-trained saliency models may provide a good initial knowledge of saliency distribution. Therefore, we conducted experiments to see if they can serve as better transmitters for transferring saliency knowledge to a new simple student model. To achieve this, starting with a set of images, we select arguably top five saliency models from the benchmark lists of state-of-of-the-art methods in public databases. Then we use these five pre-trained saliency prediction models to annotate the images selected large number of image datasets. In standard human gaze experiment, the ground-truth saliency map is calculated by recorded fixations with a Gaussian kernel to represent the saliency maps. Similar with the standard subjective experiment of eye-fixation collection, we leverage these five pre-trained saliency models to generate the saliency maps as ground-truth maps for the selected images by averaging the models predictions. We hope that this dataset can help to warm-up as a pre-training on saliency prediction field for new deep learning models without any existing pre-trained parameters. ProjectPage.
images as JPG, PNG, and etc.
Guanqun Ding*, Nevrez Imamoglu*, Ali Caglayan, Masahiro Murakawa, Ryosuke Nakamura
|First Publication Date :||Fri, 15 Apr 2022 JST|