ADAPTIVE STEREO MATCHING VIA LOOP-ERASED RANDOM WALK

This paper proposes an adaptive tree-based cost aggregation strategy for stereo matching. The previous tree-based algorithms, hindered by the greediness of minimum spanning tree (MST), provide poorly adaptive support windows and have bad performance on curved and slanted surfaces. The proposed method incorporates randomness and overcomes these drawbacks by introducing loop-erased random walk (LERW) into tree construction. Experimental results over Middlebury dataset demonstrate that our LERW-based strategy outperforms other tree-based state-of-the-art strategies in most of the high resolution test cases. Our contributions include: 1) an LERW-based cost aggregation strategy; 2) an LERW-based refinement method; 3) mathematical analysis of the adaptability of our support windows.