Label-Free Domain Adaptation for Real-Time Stereo Matching via Entropy-Guided Pseudo Labels and Epipolar Contrastive Regularization

초록

Stereo-matching networks trained exclusively on synthetic imagery suffer severe accuracy drops in realworld scenes because of domain shift. This paper introduces a three-stage, label-free adaptation pipeline that closes this gap without any target-domain depth ground truth. (i) Entropy-filtered pseudo labels supervise only the most reliable pixels, (ii) entropy minimization sharpens the entire cost volume, and (iii) an epipolar- aware contrastive learning suppresses sensitivity to color while reinforcing geometry-aware distinctiveness. Applied to FastACVNet[1], this method reduces KITTI-2015[18] D1-all and EPE, yet still runs in real time at 15 FPS on a single RTX 3090. The approach requires no image translation, extra labels, or runtime overhead, providing a plug-and-play upgrade for stereo perception in autonomous-driving and robotics applications.

키워드