Skip to content

Translation in progress Some entries still include text in the original language.

Failure-First·Indexed by this site · Originally published

Point-VLM 用几何奖励信用分配强化 3D 理解

[Daily Paper] Reinforcing 3D Understanding in Point-VLMs via Geometric Reward Credit Assignment

AI summary

论文提出几何奖励信用分配(GRCA),把 Point-VLM 的序列级广播奖励拆成按几何字段的定向监督,在基于 ShapeNetCore 的校准基准上将 3D KPA 从 0.64 提升到 0.93。各几何字段(x、y、z 坐标、边界框尺寸、旋转角)的奖励只路由到对应 token 区间,重投影一致性项惩罚投影回 2D 后与观测图像矛盾的 3D 预测,3D BBox IoU 为 0.686。

Read originalfailurefirst.org