The 31st ACM International Conference on Multimedia (ACM MM) will be held in Ottawa, Canada from October 29 to November 3, 2023. Two papers guided by Associate Professor He Tang have been accepted by ACM MM, an A-level international academic conference recommended by CCF and widely concerned in the field of multimedia computing and computer graphics. This conference received a total of 3071 valid submissions from all over the world, and finally accepted 902 papers, with an acceptance rate of about 29.3%.
Title: Partitioned Saliency Ranking with Dense Pyramid Transformers
Corresponding Author: He Tang
Co-first Authors: Chengxiao Sun (Master,School ofSoftware Engineering, 2021), Yan Xu (Master,School ofSoftware Engineering, 2021)
Project Link:https://github.com/ssecv/PSR
Abstract: This paper proposesa non-sequential method for saliency ranking, named Partitioned Saliency Ranking (PSR). It iteratively performs unordered segmentation on the most salient targets and ranks the targets based on these unordered segmentation results. By alleviating the inherent ambiguity in saliency ranking, PSR can achieve significant improvements in ranking accuracy compared to traditional methods. Additionally, a Dense Pyramid Transformers (DPT) is introduced to enable global cross-scale interactions, which significantly enhances feature interactions with the deduced computational burden. Extensive experiments demonstrate that PSR outperforms all existing methods. The source code is available athttps://github.com/ssecv/PSR.
Figure 1 (a) Ranking by Sorting and (b) Ranking by Partition
Title:Unite-Divide-Unite: Joint Boosting Trunk and Structure for High-accuracy Dichotomous Image Segmentation
Corresponding Author: He Tang
Co-first Authors: JialunPei(Postdoctoral, Chinese University of Hong Kong), Zhangjun Zhou (Master,School of Software Engineering, 2021)
Project Link: https://github.com/ssecv/UDUN
Abstract: This paper introduces a novel Unite-Divide-Unite Network (UDUN) that restructures and bipartitely arranges complementary features to boost the effectiveness of trunk and structure identification simultaneously. Thus, it can achieve high-precision segmentation of targets. UDUN can extract more detailed image features while maintaining a lightweight model, achieving 65.3 FPS even with high-resolution (1024×1024) input. UDUN outperformsall existing methods on multiple evaluation metrics. The source code is available athttp://github.com/PJLallen/UDUN.
Figure 2: Architecture of UDUN
