Bounding box matching : a sparse object-centric correspondence method for stereo vision

Abstract

In this work a simplified method for sparse, object-centric disparity estimation is proposed. It combines the state-of-the-art object detector YOLOv4 with image rectification to produce a disparity map with high speed suitable for real-time applications. Similarly, as other methods based on convolutional neural networks, this approach uses contextual and semantic image information and is robust to ill-posed image regions such as reflective, textureless and occluded regions, but requires less computational resources at the expense of detail and estimation accuracy. The method has been implemented on the Tensorflow platform and has been consequently deployed with the LabVIEW graphical programming language. It has been shown that the method works best at large distances, small object depth to distance ratios and moderate eccentricities. The source code can be found at: https://github.com/tsindel/bbox-matching.