TUVF: Learning Generalizable Texture UV Radiance Fields
Authors: An-Chieh Cheng, Xueting Li, Sifei Liu, Xiaolong Wang
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform our experiments on synthetic and real-world object datasets where we achieve not only realistic synthesis but also substantial improvements over state-of-the-arts on texture controlling and editing. |
| Researcher Affiliation | Collaboration | An-Chieh Cheng1 Xueting Li2 Sifei Liu2 Xiaolong Wang1 1UC San Diego 2NVIDIA |
| Pseudocode | Yes | Algorithm 1 : The training phase of our approach consists of two stages: (1) Canonical Surface Auto-encoder (2) Texture Feature Generator using adversarial objectives |
| Open Source Code | Yes | Codes and trained models are publicly available at https://github.com/Anjie Cheng/TUVF. |
| Open Datasets | Yes | We used 3D shapes from Shape Net s chair and car categories (Chang et al., 2015). For the 2D datasets, we employed Compcars (Yang et al., 2015) for cars and Photoshape (Park et al., 2018b) for chairs. ... For quantitative evaluation, we use 250 shapes of cats from SMAL (Zuffi et al., 2017). |
| Dataset Splits | No | For fair comparisons, we follow Texturify, splitting the 1,256 car shapes into 956 for training and 300 for testing. We apply the same split within the subset for the chair experiment, yielding 450 training and 150 testing shapes. No explicit validation split is mentioned. |
| Hardware Specification | Yes | All experiments are performed on a workstation equipped with an AMD EPYC 7542 32-Core Processor (2.90GHz) and 8 Nvidia RTX 3090 TI GPUs (24GB each). |
| Software Dependencies | Yes | We implement our framework using Py Torch 1.10. |
| Experiment Setup | Yes | During training, we sample patches starting from a minimal scale... we use larger patches (128 128)... We follow Texturify s setup in all experiments, training on 512 512 resolution images and rendering images at a resolution of 512 512 and subsequently downsampling to 256 256 for evaluation. We use an initial beta value of 1e 4 and gradually anneal it to 0.8 after processing 1e7 images. |