S23DR 2026: End-to-End 3D Wireframe Prediction via DETR-Style Set Prediction with Contrastive Denoising
Researchers have developed WireframeDETR, a novel method for predicting 3D building wireframes from multi-view point clouds, submitted to the S23DR 2026 Challenge. This approach utilizes DETR-style set prediction directly on 3D point clouds, bypassing intermediate vertex detection. Key innovations include contrastive denoising for stable training, a multi-scale encoder for feature aggregation, and progressive auxiliary loss weighting to optimize gradient flow. The model achieved a public test HSS of 0.575 and a best validation HSS of 0.534. AI
IMPACT Novel approach to 3D wireframe prediction could advance computer vision and reconstruction tasks.