Posted by spl | Publication
We propose SLTN, a Shadow and Lighting Transformation Network for efficient 3D shape recognition. Unlike prior methods that only optimize viewpoints, SLTN jointly learns camera poses and illumination, leveraging a differentiable shadow renderer to integrate cast shadows as discriminative cues. This joint optimization reduces viewpoint ambiguity and enriches geometric perception within a single rendered view. Experiments on ModelNet10/40 show that SLTN achieves single-view accuracy comparable to multi-view baselines such as MVTN with 12 views, while using fewer parameters and significantly shorter inference time. These results demonstrate the effectiveness of shadow- and lighting-aware rendering for resource-constrained 3D recognition and highlight its potential in robotics, AR/VR, and other real-time systems.