Structure from Duplicates: Neural Inverse Graphics from a Pile of Objects
Authors: Tianhang Cheng, Wei-Chiu Ma, Kaiyu Guan, Antonio Torralba, Shenlong Wang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate the efficacy of our model on a new dataset called Dup, which contains synthetic and real-world samples of duplicated objects since current multi-view datasets lack duplication samples. This allows us to benchmark inverse rendering performance under single-view or multi-view settings. |
| Researcher Affiliation | Academia | Tianhang Cheng1 Wei-Chiu Ma2 Kaiyu Guan1 Antonio Torralba2 Shenlong Wang1 1University of Illinois Urbana-Champaign 2Massachusetts Institute of Technology |
| Pseudocode | No | The paper describes its methods using prose and mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code available at https://github.com/Tianhang-Cheng/Sf D |
| Open Datasets | No | Since existing multi-view datasets do not contain duplicate objects, we collect Dup, a novel inverse rendering dataset featuring various duplicate objects. Dup consists of 13 synthetic and 6 real-world scenes... For synthetic data, we acquire 3D assets from Poly Haven and utilize Blender Cycles for physics-based rendering. As for the real-world data, we place the objects in different environments and capture 10-15 images using a mobile phone. |
| Dataset Splits | No | The paper states they use "10-300 images per scene" and "10-15 images" for real-world data, and "train our model and the baselines using 100 different views" for multi-view, but it does not specify explicit training, validation, and test dataset splits or their percentages/counts. |
| Hardware Specification | Yes | All experiments are conducted on a single Nvidia A40 GPU. |
| Software Dependencies | No | The paper mentions specific tools and models like "COLMAP", "Omnidata model", and "Blender Cycles", and optimizers like "Adam", but it does not provide specific version numbers for any of these software dependencies. |
| Experiment Setup | Yes | min S,C Ecolor + λ1Ereg + λ2Emask + λ3Enormal, (4) where λ1 = 0.1, λ2 = λ3 = 0.5. [...] min M,ω,ϕ Ecolor + λ4Esparse + λ5Esmooth + λ6Emetal, (5) where λ4 = 0.01, λ5 = 0.1, λ6 = 0.01. [...] We use the Adam optimizer with an initial learning rate 1e-4. |