stratified sampling

**Stratified sampling** is the **ray-sampling strategy that divides an interval into bins and draws samples within each bin to reduce estimator variance** - it improves coverage and training stability in volumetric rendering. **What Is Stratified sampling?** - **Definition**: Ray segments are partitioned and sampled with jitter to avoid clustering artifacts. - **Variance Control**: Even sample distribution lowers Monte Carlo variance compared with naive random sampling. - **NeRF Use**: Common in coarse rendering passes during training and inference. - **Deterministic Mode**: Can switch to fixed bin centers for reproducible evaluation. **Why Stratified sampling Matters** - **Gradient Quality**: More uniform ray coverage improves optimization signal consistency. - **Artifact Reduction**: Helps prevent missed thin structures and noisy opacity estimates. - **Efficiency**: Provides strong baseline quality without complex adaptive logic. - **Theoretical Soundness**: Well-understood estimator behavior makes tuning easier. - **Pipeline Compatibility**: Works well with hierarchical resampling and importance sampling steps. **How It Is Used in Practice** - **Bin Count**: Tune sample count per ray based on scene complexity and latency budget. - **Jitter Policy**: Use randomized jitter in training and deterministic sampling for benchmarks. - **Hybrid Setup**: Pair stratified coarse pass with fine importance pass for best tradeoff. Stratified sampling is **a standard low-variance sampling technique in NeRF pipelines** - stratified sampling remains a reliable default when balancing rendering quality and computational cost.