[2603.26258] ARTA: Adaptive Mixed-Resolution Token Allocation for Efficient Dense Feature Extraction

Computer Science > Computer Vision and Pattern Recognition arXiv:2603.26258 (cs) [Submitted on 27 Mar 2026] Title:ARTA: Adaptive Mixed-Resolution Token Allocation for Efficient Dense Feature Extraction Authors:David Hagerman, Roman Naeem, Erik Brorsson, Fredrik Kahl, Lennart Svensson View a PDF of the paper titled ARTA: Adaptive Mixed-Resolution Token Allocation for Efficient Dense Feature Extraction, by David Hagerman and 4 other authors View PDF HTML (experimental) Abstract:We present ARTA, a mixed-resolution coarse-to-fine vision transformer for efficient dense feature extraction. Unlike models that begin with dense high-resolution (fine) tokens, ARTA starts with low-resolution (coarse) tokens and uses a lightweight allocator to predict which regions require more fine tokens. The allocator iteratively predicts a semantic (class) boundary score and allocates additional tokens to patches above a low threshold, concentrating token density near boundaries while maintaining high sensitivity to weak boundary evidence. This targeted allocation encourages tokens to represent a single semantic class rather than a mixture of classes. Mixed-resolution attention enables interaction between coarse and fine tokens, focusing computation on semantically complex areas while avoiding redundant processing in homogeneous regions. Experiments demonstrate that ARTA achieves state-of-the-art results on ADE20K and COCO-Stuff with substantially fewer FLOPs, and delivers competitive performance ...