ARS: Enhancing Depth Estimation in Multi-View 3D Detection via Adaptive Ray Sampling
Adaptive Ray Sampling improves multi-view 3D detection by dynamically sampling hard negatives along rays to reduce depth ambiguity during training, boosting precision by 3% on nuScenes without inference overhead.
Published Jun 6, 2025Paper ↗
Only vote on papers you've read. Sign in to vote.
Abstract
Multi-view 3D object detection remains challenged by inaccurate depth estimation, leading to duplicate detections and reduced precision. Existing approaches often suffer from rigid sampling mechanisms and an inability to adapt to varying scene complexities, which amplifies depth ambiguities and further compromises detection accuracy. Inspired by RayDN's ray denoising strategy, we propose Adaptive Ray Sampling (ARS), a novel training technique that dynamically constructs an adaptive sampling distribution across training stages. ARS selectively samples hard negative instances along rays, improving depth estimation and enhancing detection precision. As a framework-agnostic, plug-and-play module, ARS seamlessly integrates with any DETR-style model, operating solely during training without affecting inference efficiency. Experiments on the nuScenes validation set demonstrate that ARS improves precision by 3% over StreamPETR, highlighting its effectiveness in mitigating depth estimation errors.