SuperFormer: Superpixel-based Transformers for Salient Object Detection

Jinman Park, Zahra Gharaee, Paul W. Fieguth

Keywords: Salient Object Detection, Superpixels, Transformers, Graph Neural Networks

TL;DR: Salient Object Detection using superpixel representations and Transformers

Abstract: Images often have local redundant information that can strain the training of deep neural networks. An effective way to reduce spatial redundancy and image complexity is to over-segment with superpixels. With a fast, linear computational complexity, Simple Linear Iterative Clustering (SLIC) generates superpixels by grouping pixels as a function of colour similarity and spatial proximity. However, it is challenging and non-trivial to train a model on over-segmented images with dynamic graph structure and low spatial inductive bias. In order to train on unstructured data, graph neural networks (GNNs) can be applied to classify each superpixel for salient object detection (SOD) by considering a set of superpixels as graphs. Although other works on graph classification or node classification were able to utilize pre-defined edge information or GNNs, naive applications on superpixel graphs do not translate trivially. Our proposed SuperFormer method introduces new feature attributes for superpixels and a dynamic positional encoding for heterogeneous spatial graphs to achieve state-of-the-art results in salient object detection for low model complexity.

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6560