Sharf: Shape-conditioned Radiance Fields from a Single View
Authors: Konstantinos Rematas, Ricardo Martin-Brualla, Vittorio Ferrari
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate in several experiments the effectiveness of our approach in both synthetic and real images. Section 5. Experiments: We extensively evaluate our approach for novel view synthesis given a single image on three datasets: (1) an existing benchmark based on Shape Net (Shape Net-SRN (Sitzmann et al., 2019), Sec. 5.2); (2) an updated Shape Net test set with images rendered more realistically and at higher resolution (dubbed Shape Net-Realistic, Sec. 5.3); (3) the Pix3D (Sun et al., 2018) dataset, with photographs of real objects (Sec. 5.4). Metrics. We evaluate with the standard image quality metrics PSNR and SSIM (Zhou Wang et al., 2004). |
| Researcher Affiliation | Industry | 1Google Research. Correspondence to: Konstantinos Rematas <krematas@google.com>. |
| Pseudocode | No | The paper describes methods in prose and with diagrams, but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any statements about open-sourcing code or provide links to a code repository. |
| Open Datasets | Yes | We use the same experimental setup and data as (Sitzmann et al., 2019). The dataset consist of 6591 Chairs and 3514 Cars that are split for training, validation and testing. |
| Dataset Splits | No | The paper states, 'The dataset consist of 6591 Chairs and 3514 Cars that are split for training, validation and testing.' However, it does not provide the specific percentages or exact sample counts for each split, deferring to the referenced paper for the 'experimental setup and data'. |
| Hardware Specification | No | The paper does not mention any specific hardware used for running the experiments (e.g., GPU models, CPU types, or cloud computing instances). |
| Software Dependencies | No | The paper mentions 'Adam (Kingma & Ba, 2015)' as an optimizer but does not specify version numbers for any key software components or libraries (e.g., Python, PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | We use 128 rays per image and 128 stratiļ¬ed samples per ray, plus 128 importance samples estimated by the geometric scaffold. The latent codes are initialized from a normal distribution and all optimizations are performed with Adam (Kingma & Ba, 2015). During inference, we optimize the latent codes and networks for 1000 iterations. |