Neural Radiance Field Codebooks
Authors: Matthew Wallingford, Aditya Kusupati, Alex Fang, Vivek Ramanujan, Aniruddha Kembhavi, Roozbeh Mottaghi, Ali Farhadi
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our decomposition and representations on several downstream tasks: unsupervised segmentation (real and synthetic), object navigation, and depth ordering. NRC shows improvement over baseline methods on all three tasks. |
| Researcher Affiliation | Collaboration | Matthew Wallingford1, Aditya Kusupati1, Alex Fang1, Vivek Ramanujan1, Aniruddha Kembhavi2, Roozbeh Mottaghi1, Ali Farhadi1 1University of Washington; 2PRIOR, Allen Institute for AI |
| Pseudocode | No | The paper describes the method using text, diagrams (Figure 2), and mathematical equations, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | No | The paper mentions 'We use the implementation available at https://github.com/KovenYu/uORF' which refers to a baseline method, not the code for NRC itself. There is no explicit statement or link indicating that the source code for NRC is open-sourced or available. |
| Open Datasets | Yes | Proc THOR (Deitke et al., 2022) consists of procedurally generated indoor scenes similar to Robo THOR. CLEVR-3D (Johnson et al., 2017) is a synthetic dataset consisting of geometric primitives from multiple views and is used for unsupervised segmentation. The NYU Depth Dataset (Silberman et al., 2012) consists of images from real-world indoor scenes accompanied by depth and segmentation maps. |
| Dataset Splits | Yes | Robo THOR consists of 89 indoor scenes split between train, validation, and test. Following the convention of Stelzner et al. (2021), we test on the first 500 scenes of the validation set and report foreground-adjusted random index (FG-ARI). |
| Hardware Specification | Yes | We train models on a single NVIDIA A40. |
| Software Dependencies | No | The paper mentions architectural components (e.g., ResNet34, ResNet50), optimizers (ADAM), and specific algorithms (DD-PPO), but does not provide specific version numbers for software libraries or dependencies (e.g., PyTorch version, CUDA version). |
| Experiment Setup | Yes | We train with a learning rate of 1e-4 with the ADAM optimizer. We set the near field to .4 and far field to 5.5. We train from scratch, and from the video Mo Co model trained on Proc THOR. We found starting from the pretrained initialization reduced the number of training epochs required for convergence. For the model from scratch we train for 300 epochs. For the model from Mo Co initialization we train for 100 epochs. While training we use a sliding window of 5 frames to determine corresponding images in Proc THOR. We train a policy using DD-PPO for 200M steps. |