Delaunay-Guided Adaptive Superpixel Coding
PhD. Seminar by
Mustafa Mohammadi Gharasuie
Date: April 8th, 2026
Time: 3:00 PM – 5:00 PM
Location: 354 Dillon Hall
Abstract:
Self-supervised visual representation learning has achieved strong transfer performance, but modern Transformer-based pipelines remain computationally expensive due to dense token interactions that scale quadratically with the number of tokens. Adaptive Superpixel Coding (ASC) partially alleviates this issue by grouping tokens via affinity graphs and connected components, yet it still relies on dense all-pairs similarity computations that become costly at high resolutions. We propose Delaunay-Guided Adaptive Superpixel Coding (D-ASC), a sparse, geometry-aware alternative that replaces grid patches with superpixel tokens and dense interactions with a bounded-degree Delaunay graph prior. D-ASC restricts self-attention to local graph neighborhoods, augmented with a small set of global tokens for limited long-range interaction, and performs prior-aware grouping using sparse edge scoring followed by connected components on the sparse graph. D-ASC preserves ASC’s adaptive merging behavior while significantly improving memory efficiency and training and speed. We describe the method, analyze its complexity, and evaluate representation transfer on standard downstream tasks.
PhD Doctoral Committee:
Internal Reader: Prof. Alioune Ngom
Internal Reader: Prof. Mehdi Sangani Monfared
External Reader: Prof. Dan Wu
Advisor (s): Prof. Luis Rueda
