NEURAL RADIANCE FIELDS INNOVATION

Theoretical Basis & Color Space Model

The Bi-Illuminant Dichromatic Reflection (BIDR) model highlights how a logarithmic transformation can simplify the separation of illumination and reflectance. In linear RGB, material and illumination terms are conflated, requiring complex representations (two colors per material, varying lengths, curves for sRGB).

In contrast, log RGB space separates these terms: log I = log RB + log(A+YD). This allows for a more compact representation, where each material is defined by a single point RB, and illumination effects become additive offsets or vectors. This theoretical advantage suggests log RGB could enable NeRF to learn a significantly more efficient and robust representation of scene appearance.

We introduce TrueLog, a normalized logarithmic transformation that maps linear signals from [0, 1] to [0, 1] in log space, ensuring mathematical stability and constraining network outputs. This contrasts with GoPro's GPLog and the gamma-corrected sRGB which exhibit less ideal properties for scene representation.

Experimental Methodology

To rigorously test our hypothesis, we captured approximately 30 videos using a GoPro Hero 13 Black, specifically utilizing its 10-bit GPLog encoding to allow for accurate linearization to RGB. Camera poses were extracted using COLMAP, with challenging dark scenes requiring conversion to TrueLog for successful registration.

We used the BiLaRF NeRF model [31] as our base architecture and introduced crucial modifications: an implicit-to-linear transform (A), a linear-to-sRGB transform (B) for loss computation, and an input data transform (C) to ensure fair comparisons. This setup allowed us to train NeRFs in different internal representation spaces—GPLog, linear RGB, sRGB, and TrueLog—while ensuring consistent sRGB loss calculation.

Experiments were conducted across various conditions, including different network sizes (MLP width, grid encoder size), training durations (up to 25,000 epochs), and multiple runs for robustness analysis, all supervised with sRGB data and Charbonnier loss.

Performance & Results

Our comprehensive evaluation consistently shows that training NeRF models in TrueLog space significantly outperforms other color spaces across multiple metrics. TrueLog exhibits higher PSNR, superior robustness (lower standard deviation in repeated runs), and remarkable performance in challenging low-light conditions.

TrueLog's advantage is most pronounced in low-light scenarios (e.g., GX010032), where it recovers significantly more scene detail and produces sharper images and more accurate depth maps (Figure 2, Figure 3). This robustness extends to network compactness, with TrueLog maintaining performance better than other spaces even with reduced MLP widths and grid encoder sizes (Figure 6).

Strategic Implications

For enterprises leveraging NeRF for digital twins, virtual reality, or advanced visualization, adopting a logarithmic representation like TrueLog offers a strategic advantage. It leads to more accurate and robust 3D models, especially when dealing with varied or challenging real-world lighting conditions, and improves computational efficiency by reaching high quality results with fewer training iterations.

This work demonstrates that the choice of color space is not a minor detail but a critical factor influencing NeRF performance, stability, and data efficiency. The findings generalize beyond the specific BiLaRF model, showing consistent benefits across different NeRF approaches like Mip-NeRF and Robust-NeRF (Figure 7).

Future work will explore the generalization of log-space representations to a broader range of NeRF architectures, including dynamic scenes and different tasks, further solidifying its role in advanced neural rendering.