photometric loss

**Photometric loss** is the **objective that measures color differences between rendered predictions and reference images at sampled pixels** - it is the primary supervision signal in many neural rendering pipelines. **What Is Photometric loss?** - **Definition**: Compares predicted RGB values to ground truth using L1, L2, or robust variants. - **Application**: Used at ray-sampled pixels during NeRF and view-synthesis training. - **Sensitivity**: Affected by exposure changes, motion blur, and pose misalignment. - **Extensions**: Often combined with perceptual, depth, or regularization losses for better stability. **Why Photometric loss Matters** - **Core Supervision**: Directly drives reconstruction quality in learned scene representations. - **Optimization Signal**: Strong photometric gradients help recover geometry and appearance jointly. - **Metric Alignment**: Correlates with PSNR-style image-fidelity reporting. - **Failure Diagnosis**: Loss plateaus can indicate calibration or sampling issues. - **Limitations**: Alone it may not enforce temporal or geometric consistency in dynamic settings. **How It Is Used in Practice** - **Robust Variant**: Use Charbonnier or Huber style losses for outlier resilience. - **Color Handling**: Normalize color space and exposure to reduce supervision noise. - **Loss Balancing**: Weight photometric loss with geometry priors for stable convergence. Photometric loss is **the baseline reconstruction objective in neural view synthesis** - photometric loss works best when paired with calibration hygiene and complementary structural constraints.