RSD-BEV: Residual Self-distillation Framework for Efficient BEV Representation Learning

초록

Bird’s-eye-view (BEV) representation has become a standard for autonomous driving and robotics applications. However, due to the weak inductive bias and inherent geometric ambiguity in view transformation, BEV models are data-hungry, making data-efficient training a critical research challenge. To address this, recent works have explored self-distillation strategies that leverage the model’s own predictions to provide dense supervisory signals. While diffusion-based self-distillation has shown promise, it struggles with high computational overhead caused by iterative denoising on high-dimensional feature maps. To address this, we propose a residual self-distillation framework that emulates diffusion-style refinement as additive residual corrections. Our method effectively replaces the computationally expensive multi-step diffusion process with a single-step residual mapping. Experiments on the nuScenes dataset demonstrate that our approach significantly reduces computation costs while achieving performance comparable to state-of-the-art diffusion-based methods.

키워드