高德发布无长程依赖的万帧级3D重建模型ABot-Recon
高德推出ABot-Recon，用局部12帧实现万帧级流式3D重建，解决长序列误差累积和内存膨胀，达到SOTA。
高德发布ABot-Recon，一种流式3D重建模型（用连续视频流实时构建三维场景）。它摒弃长程记忆，只看最近12帧即可持续重建上万帧场景，计算量不随序列增长，在多个基准上误差最低。
正文摘录
- title: 高德发布首个无长程依赖的万帧级流式 3D 重建模型 ABot-Recon，以 12 帧重建万帧 3D 场景 - sourcecompany: 量子位 - bodymarkdown: On August 28, Amap, under Alibaba Group, officially released ABot-Recon, the first ten-thousand-frame-level streaming 3D reconstruction model without long-range dependency. On August 28, Amap, under Alibaba Group, officially released ABot-Recon, the first ten-thousand-frame-level streaming 3D reconstruction model without long-range dependency. The model requires no long-range memory — using only 12 consecutive local frames, it can reconstruct over ten thousand frames of 3D scenes in real time and stably, achieving "reconstruct while capturing." On authoritative benchmarks including KITTI, Oxford Spires, and VBR, ABot-Recon achieves SOTA with the lowest errors. As autonomous driving and embod…