Paper Reading

AdaNeRF: Adaptive Sampling for Real-time Rendering of Neural Radiance Fields [ECCV22]

NeuMesh: Learning Disentangled Neural Mesh-based Implicit Field for Geometry and Texture Editing [ECCV22 Oral]

1. 训练NeuS用来重建mesh

2. 用来网络蒸馏来训练NeuMesh：

   Although such vertex-bounded and geometry-texture disentangled representation merits good flexibility on editing purpose, it does not preserve spatial continuity as MLP-based methods [27,55,65] and thus easily suffers from unstable training. To mitigate this problem, we employ a distillation and fine-tuning training scheme, which leverages a pre-trained implicit field to guide the optimization of our representation.

3. 训练之后，颜色code和纹理code都在mesh的顶点上

NeuS NeurIPS 2021 Spotlight: [PDF] [Project Page]

TopoSeg: Topology-aware Segmentation for Point Clouds [IJCAI-22]

在点云上定义了一个拓扑loss

NeuralBF: Neural Bilateral Filtering for Top-down Instance Segmentation on Point Clouds [Arxiv]

用语义分割+convex hulls来实例分割。