https://github.com/ActiveVisionLab/Awesome-LLM-3D

https://aicarrier.feishu.cn/wiki/GvibwrWumiYVxYk3Ik5c7Yd3nSb

pipeline整理：

1. 大部分论文的pipeline：Encoder → Projector → LLM
   1. Encoder：对point clouds提取特征
   2. Projector：使用projector将特征投影到LLM所需的point token
   3. LLM：使用LLM预测token
2. Encoder → LLM
   1. An Embodied Generalist Agent in 3D World
3. Building object set → LLM
   1. ConceptGraphs

Point Encoder的variants

• 将2D image feature反投影到3D point clouds上，获得3D point clouds的feature
• Object-level encoder: I2P-MAE
• Object-level encoder: Recon++
• Object-level encoder: Point-BERT
• Object-level encoder: Ulip2 and Uni3D
• Object-level encoder: Vanilla transformer structurally equivalent to ViT + 2D model初始化
• Scene-level encoder: EPCL
• Scene-level encoder: Masked transformer encoder
• Scene-level encoder: PointNet++加上spatial transformer
• Scene-level encoder: Multi-view transformer + Fusion transformer