Spider: A Unified Framework for Context-dependent Concept Understanding

• DOI: 10.48550/arxiv.2405.01002
• Publication Date: 2024-05-02
• Authors: Xiaoqi Zhao; Youwei Pang; Wei Ji; Baicheng Sheng; Jiaming Zuo; Lihe Zhang; Huchuan Lu
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition; Machine Learning (cs.LG)

Different from the context-independent (CI) concepts such as human, car, and airplane, context-dependent (CD) require higher visual understanding ability, camouflaged object medical lesion. Despite rapid advance of many CD tasks in respective branches, isolated evolution leads to their limited cross-domain generalisation repetitive technique innovation. Since there is a strong coupling relationship between foreground background context tasks, existing methods train separate models focused domains. This restricts real-world concept towards artificial general intelligence (AGI). We propose unified model with single set parameters, Spider, which only needs be trained once. With help proposed filter driven by image-mask group prompt, Spider able understand distinguish diverse accurately capture Prompter's intention. Without bells whistles, significantly outperforms state-of-the-art specialized 8 different segmentation including 4 natural scenes (salient, camouflaged, transparent objects shadow) lesions (COVID-19, polyp, breast, skin lesion color colonoscopy, CT, ultrasound, dermoscopy modalities). Besides, shows obvious advantages continuous learning. It can easily complete training new fine-tuning parameters less than 1\% bring tolerable performance degradation 5\% for all old tasks. The source code will publicly available at \href{https://github.com/Xiaoqi-Zhao-DLUT/Spider-UniCDSeg}{Spider-UniCDSeg}.