NILUT: Conditional Neural Implicit 3D Lookup Tables for Image Enhancement
Authors: Marcos V. Conde, Javier Vazquez-Corral, Michael S. Brown, Radu Timofte
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments Learning a complete 3D LUT transformation of the 8-bit RGB space requires 2563 = 16.78 million input (and output) colors. ... Table 1 presents our results for different configurations of MLPs -a basic MLP, SIREN, and a Residual MLP-, under two different numbers of neurons (N) and layers (L). ... Ablation Studies We studied the results for a larger configuration of MLPs under the first evaluation scenario. Results are presented in Table 2. |
| Researcher Affiliation | Academia | 1 Computer Vision Lab, CAIDAS, University of W urzburg 2 Computer Vision Center (CVC) 3 Department of Computer Science, Universitat Aut onoma de Barcelona 4 York University |
| Pseudocode | No | The paper does not include pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | Code at https://github.com/mv-lab/nilut |
| Open Datasets | No | To learn our continual representations Φ we minimize: i Φ(xi) ϕ(xi) 1, where ϕ is the real 3D LUT -see Eq. 1-. This function Φ is an implicit neural representation of a 3D lookup table ϕ and can be formulated as: ... We represent the RGB space as a set X = {xi} of color pixels xi = (ri, gi, bi). This set contains 16 million elements if we consider the complete RGB space (i.e. 2563). ... Note that using this setup we do not require natural images to learn real 3D LUTs, just the corresponding RGB maps (Halds). ... The paper describes creating a training set from an RGB map (M) and 3D LUTs, but does not provide concrete access (link, citation for their specific processed version) to this training data. While MIT5K is public, it's used for testing/evaluation, not direct training. |
| Dataset Splits | No | The paper does not explicitly mention using a validation set or describe a validation process. |
| Hardware Specification | No | Table 4: CNILUT deployment on two mid-level smartphone GPUs. We report the average image processing runtime over 10 runs the std. deviation (see Figure 10). Input res. Mali-G77 (ms) Adreno 620 (ms) ... The paper mentions Mali-G77 and Adreno 620 for deployment benchmarks, but does not specify the hardware used for training the models or conducting the main experiments presented in Tables 1-3. |
| Software Dependencies | No | many operations, including trilinear interpolation on CUDA (Zeng et al. 2020) are not supported by Py Torch Mobile or TFLite, the most common frameworks for developing efficient mobile models. The paper mentions software frameworks like PyTorch Mobile and TFLite, but does not provide specific version numbers for these or any other software components used in the experiments. |
| Experiment Setup | Yes | In our experiments, we set to three and five the number of 3D LUTs to learn using this approach. ... We train our CNILUT for 10000 steps. |