Metalwork: A Synthetic Dataset and Baseline for Stereo Matching of Metal Workpieces

Stereo matching has received significant progress in recent years. However, traditional methods designed for conventional scenarios are unsuitable for metal workpieces due to challenges such as metal reflections, lack of distinct textures, and small holes. To address the challenge of stereo matching of metal workpieces, we propose (1) a large-scale synthetic dataset named MetalWork, and (2) a novel stereo matching framework for metal workpieces. Our MetalWork is the first synthetic dataset specifically designed for metal workpieces, offering a more realistic simulation of industrial challenges to support the training of stereo matching frameworks better. Meanwhile, we introduce a transformer-based context network that leverages multi-scale, multi-head attention mechanisms to effectively exchange features across both spatial and channel dimensions, enabling more accurate capture of intricate structural details. Extensive experiments show that, with the aid of MetalWork, our proposed method outperforms existing baselines in stereo matching of metal workpieces and achieves superior accuracy and robustness in disparity estimation for real metal workpieces.