New paper uses design rules to estimate 3D dimensions from a single image

A new paper from researchers Hidenori Sakaniwa, Akihito Akai, and Akihiko Hyodo introduces a knowledge-driven approach to estimate physical dimensions from a single monocular image, enabling the generation of accurate 3D assets for digital twin construction. The method targets a key problem in autonomous driving simulation: verifying in-vehicle camera performance in virtual environments often fails because 3D objects have incorrect scale compared to real-world objects. This is especially challenging for traffic signs installed at high altitude, where LiDAR or stereo cameras cannot reliably measure distance. The proposed technique decomposes an object into multiple structural elements (e.g., sign face, pole, base) and then estimates each component's size by applying external knowledge—such as standard design rules, geometric relationships between parts, and conventional dimensional constraints. This allows the system to recover absolute scale from a single image without requiring depth sensors or multiple views.

The resulting 3D assets can be placed into a digital twin environment with a scale approximating the real environment, dramatically improving the accuracy of virtual verification for autonomous driving cameras. The paper includes a 6-page description with 4 figures, demonstrating the decomposition and estimation pipeline. By bridging the gap between virtual simulations and physical reality, this work has the potential to reduce false detections and missed detections in pre-deployment testing of self-driving systems. The researchers have submitted the paper as arXiv:2606.30896 and it is currently under review. This approach could extend beyond traffic signs to any object with known structural rules, opening the door to scalable digital twin creation from single photographs.