ECCV 2026 Paper LeAD-M3D Achieves Real-Time Monocular 3D Detection Without LiDAR
rsasaki0109 · x · 2026-08-12
A paper accepted to ECCV 2026 introduces LeAD-M3D, a monocular 3D object detector that achieves state-of-the-art accuracy and real-time inference without extra modalities like LiDAR or stereo vision.
The architecture features three key components:
- A2D2 (Asymmetric Augmentation Denoising Distillation): Transfers geometric knowledge from a clean-image teacher to a MixUp-noised student via a quality- and importance-weighted depth-feature loss.
- CM3D (3D-aware Consistent Matching): Integrates 3D MGIoU into the matching score to improve prediction-to-ground truth assignment for stable supervision.
- CGI3D (Confidence-Gated 3D Inference): Accelerates inference by restricting expensive 3D regression to confident regions.
Experiments show LeAD-M3D sets a new Pareto frontier, achieving SOTA 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
- Report: Ilya Sutskever's SSI Pivots to Test-Time Training for New Reasoning Engine — iruletheworldmo · 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
- Turing Institute Research Identifies Critical Gaps in Cyberattack Data — turinginst · 2026-08-12
- Paper Decodes Hidden Reasoning Traces from Claude and GPT — Zealousideal_Sort74 · 2026-08-12
- Empirical Test: Are LLMs Actually Confident During First Factual Recall? — Any-Chipmunk5480 · 2026-08-12