STUR3D：面向3D目标检测的时空统一表征学习
STUR3D: Spatio-Temporal Unified Representation Learning for 3D Object Detection
把历史时序3D特征投影回2D图像平面来监督/蒸馏2D检测器，并反向把时序线索注入2D检测以生成3D查询，这套跨维度对齐思路可直接迁移到BEV/环视检测器的查询初始化与遮挡场景下的时序建模。
原文依据 we project temporal 3D features to the 2D image plane, empowering the 2D detector to distill representations essential for 3D localization
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论现有的环视3D目标检测器利用当前2D信息初始化高置信度查询，同时将历史3D特征作为先验。然而，这种对2D线索的严重依赖会在2D与3D表征之间引入时空不一致性。具体而言，2D线索缺乏足够的空间信息，限制了3D定位能力；此外，时序交互不足往往导致遮挡情况下的目标漏检。为应对这些挑战，我们提出STUR3D，一个在2D与3D感知之间建立时空对齐的统一框架。首先，我们将时序3D特征投影到2D图像平面，使2D检测器能够蒸馏出对3D定位至关重要的表征，从而协调跨维度信息。其次，我们将时序线索注入2D检测，促进时空推理，确保在动态场景与遮挡条件下实现鲁棒的3D检测。此外，我们将深度感知的几何线索嵌入特征以用于2D到3D的提升，缓解固有的歧义。大量nuScenes实验验证了STUR3D，在测试集上以57.9%的平均精度均值和64.6%的NDS取得SOTA。
英文原文摘要 1179 字符
Existing surrounding-view 3D object detectors initialize high-confidence queries using current 2D information, while leveraging historical 3D features as priors. However, such heavy reliance on 2D cues introduces spatio-temporal inconsistencies between 2D and 3D representations. Specifically, 2D cues lack sufficient spatial information, limiting 3D localization capability. Moreover, insufficient temporal interaction often leads to object omission under occlusion. To address these challenges, we propose STUR3D, a unified framework establishing spatio-temporal alignment between 2D and 3D perception. First, we project temporal 3D features to the 2D image plane, empowering the 2D detector to distill representations essential for 3D localization, harmonizing cross-dimensional information. Second, we inject temporal cues into 2D detection, fostering spatio-temporal reasoning, and ensuring robust 3D detection under dynamic scenes and occlusion. Additionally, we embed depth-aware geometric cues into features for 2D-to-3D lifting, mitigating inherent ambiguities. Extensive nuScenes experiments validate STUR3D, achieving SOTA on the test set with 57.9% mAP and 64.6% NDS.