LeAD-M3D: LiDAR-Free Monocular 3D Detection Achieves SOTA at 3.6x Speedup
rsasaki0109 · x · 2026-08-12
LeAD-M3D introduces a real-time monocular 3D object detector that operates without LiDAR, stereo vision, or strong geometric priors.
The method achieves a new Pareto frontier through three key components:
- A2D2 (Asymmetric Augmentation Denoising Distillation): Distills geometric knowledge from a clean-image teacher to a MixUp-noised student.
- CM3D (3D-aware Consistent Matching): Integrates 3D MGIoU into matching scores for stable and precise supervision.
- CGI3D (Confidence-Gated 3D Inference): Accelerates inference by restricting expensive 3D regression to confident regions.
LeAD-M3D achieves state-of-the-art accuracy on KITTI and Waymo, and the best reported car AP on Rope3D, while running up to 3.6× faster than prior high-accuracy models like MonoDiff. Code and weights will be released.
Related event: LeAD-M3D Achieves LiDAR-Free Real-Time Monocular 3D Detection(2 posts)→
More from Research
- Paper Proves No Gradient Descent Stepsize Schedule Can Match Nesterov Acceleration — prof_grimmer · 2026-08-12
- Roboflow's Open-Source Trackers Library Adds McByte for Occlusion Handling — burny_tech · 2026-08-12
- Report: Ilya Sutskever's SSI Pivots to Test-Time Training for New Reasoning Engine — iruletheworldmo · 2026-08-12
- 300+ Real-World ML System Design Case Studies from 80+ Top Companies — mdancho84 · 2026-08-12
- Fine-tuning Muse Glimmer 30B Boosts Click Grounding Accuracy to 41% — mervenoyann · 2026-08-12
- Liquid Crow Released: Squeezing Physical Cognition into a 450M-Parameter Micro-World Model — helloiamleonie · 2026-08-12