Look Where You Can: Active View Selection for CAD Reconstruction under Occlusion
The research introduces SightCAD, a framework for parametric CAD reconstruction that solves the problem of limited visual access to real-world objects. It uses a learned view selector to choose $K$ feasible views from all possible angles for a vision-language model (VLM). This joint training approach outperforms traditional selection methods like SA-max by up to 6.4 IoU points, leading to superior CAD reconstruction quality across various complex benchmarks.