Geometry-Aware Projective Mapping for Unbounded Neural Radiance Fields

Authors: Junoh Lee, Hyunjun Jung, Jin-Hwi Park, Inhwan Bae, Hae-Gon Jeon

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To demonstrate the effectiveness of the proposed mapping function and ray parameterization, we conduct extensive experiments on unbounded scenes, including 360 object-centric and free trajectory. Experimental results show that our method can be successfully integrated with existing multilayer perceptron (MLP)-based and voxel-based models and contribute to significant performance improvements for novel view synthesis, where the conventional mapping functions often fail.
Researcher Affiliation Academia Gwangju Institute of Science and Technology, Gwangju, Korea
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper mentions using and customizing public codes for comparison methods (e.g., DVGO, Tenso RF, Ne RF, iNGP re-implementation) but does not provide a statement or link for the open-source code of its own methodology.
Open Datasets Yes For our experiments, we use three datasets: two 360 object-centric datasets Tanks and Temples Zhang et al. (2020) and mip-Ne RF 360 Barron et al. (2022) and a free trajectory dataset, named Free Dataset Wang et al. (2023).
Dataset Splits No The paper mentions training steps and iterations but does not provide specific details on how the datasets were split into training, validation, or test sets (e.g., exact percentages, sample counts, or explicit references to predefined splits with citations for reproducibility). It mentions following the training configurations of other frameworks but not their specific data splits.
Hardware Specification No The paper mentions running experiments and training models but does not provide specific details about the hardware used, such as GPU or CPU models, memory, or cloud computing instance types.
Software Dependencies No The paper mentions using the "Py Torch framework" and integrating with "DVGO", "Tenso RF", "i NGP", and "Ne RF" models. However, it does not provide specific version numbers for any of these software dependencies.
Experiment Setup Yes To implement the DVGO baseline, we adopt a voxel model from Sun et al. (2022); Cheng et al. (2022). The number of voxels is initially 503, and we gradually scale up the voxel grid at the [2, 000, 4, 000, 6, 000, 8, 000, 10, 000, 12, 000, 14, 000, 16, 000] training steps up to 3203 voxels. The feature dimension of the voxel grid is set to 12. For the viewing direction d, we embed it within a positional embedding. There are two hidden layers with 128 channels in the shallow MLP layer. We skip the low-density query points in the unknown space using the threshold 10 4. The point sampling is performed using the marching step strategy; input points on a ray are moved in small steps, which is set to half of the voxel sizes. Since the size of the marching step in real-world space and embedding space may be different, it samples more points on the ray and prunes the oversampled points. We optimize scene representations using the Adam optimizer with a batch size of 213 rays for 40k iterations. All voxel grids base learning rates are 5 10 4, while the MLP s learning rate is 10 3. Applying the exponential learning rate decay, the learning rates are scaled down by 0.1 after 20k iterations.