CoFie: Learning Compact Neural Surface Representations with Coordinate Fields
Authors: Hanwen Jiang, Haitao Yang, Georgios Pavlakos, Qixing Huang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This paper introduces Co Fie, a novel local geometry-aware neural surface representation. When using the same amount of parameters with prior works, Co Fie reduces the shape error by 48% and 56% on novel instances of both training and unseen shape categories. Moreover, Co Fie demonstrates comparable performance to prior works when using even 70% fewer parameters. Code and model can be found here: https://hwjiang1510.github.io/Co Fie/Experiments: Section 5. |
| Researcher Affiliation | Academia | Hanwen Jiang Haitao Yang Georgios Pavlakos Qixing Huang Department of Computer Science, The University of Texas at Austin {hwjiang,yanght,pavlakos,huangqx}@cs.utexas.edu |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code and model can be found here: https://hwjiang1510.github.io/Co Fie/ |
| Open Datasets | Yes | We train Co Fie on 1000 shape instances sample from Shape Net [6] of chairs, planes, tables, lamps, and sofas (200 instances for each category). |
| Dataset Splits | No | The paper does not explicitly provide training/test/validation dataset splits or percentages. It mentions training with 1000 shape instances and testing with specific test sets but no explicit split information. |
| Hardware Specification | Yes | We train on 4 GPUs with 24GB memory for 1 day. |
| Software Dependencies | No | The paper mentions using Adam optimizer [21] but does not provide specific version numbers for any software libraries, frameworks (like PyTorch or TensorFlow), or programming languages. |
| Experiment Setup | Yes | We use latent code of size 125 for all cells. The MLP is composed of 5 layers where the first 4 layers are linear layers and the last layer is quadratic. The hidden channel size is 128. We use the voxel grid size of 32 32 32. During training, we use 12 shapes for each batch. For each shape, we sample 3000 voxels that intersect with the surface of the shape (with return). We sample 24 points for each cell for training, and each point is sampled within 1.5 times the radius of the voxel to ensure boundary consistency between cells. We use the Adam optimizer [21] with learning rates 5e 4, 1e 3, and 1e 3 for the MLP, coordinate fields, and latent codes. We train with 150000 iterations and reduce the learning rates by half for every 20000 iteration. During inference, we use a learning rate of 5e 4 for 800 iterations. Reconstructed meshes are obtained by performing Marching Cubes with a 128 resolution by default. We use the quaternion representation for the rotation matrix of the coordinate frames. |