Open code, practice, and notes →
Download source with English comments · for reading
Gaussian Splatting · 21 Defining Quality Evaluation Criteria
Series contents · Technology map
Stage: Visibility and Quality Experiments · Editable example · Storage ID: gaussian-splatting-step-21
Goals and learning outcomes
- Fix the comparison reference.
- Interpret numerical and visual evidence together.
Prerequisites
Continue using the output and records from Unit 20. Before the Defining Quality Evaluation Criteria experiment, preserve the previous inputs and reference outputs.
Key concepts and explanations
Reference image
Interpretation depends on whether the reference is synthetic or an actual observed photograph. Do not call image differences calculated without ground-truth photographs reconstruction accuracy.
Multiple observations
A single average error can hide thin structures or flicker. Evaluate specified regions together with a camera path.
Practice setup
First check the execution scope guidance on the example screen. The following describes the environment needed to extend the assignment in a separate project. This is a document-guided exercise in a separate local project. Use only small data you create or local files whose usage permission you have checked. The existing preview is a CPU teaching viewer for .splat/ASCII GS PLY files of at most 1 MiB and 512 items; it is not a Spark GPU renderer.
The linked example is a starting point. It does not implement every task below. Complete work beyond its supported scope in a separate project.
Practice tasks
- In evaluation.json, define three viewpoints, two regions of interest, and a background.
- Save identically sized ImageData for baseline and reduction modes.
- In report.md, write the mean absolute channel difference and describe losses in the regions of interest.
Variables to observe
| Variable | Baseline | Change | Expected change / hypothesis |
|---|---|---|---|
| Evaluation region | Full screen | Thin-structure ROI | Expect local losses to stand out more than in the overall average. |
Predictions in this table are not measured results. Change one condition at a time and record why the actual result differs.
Completion criteria
- Use the same ROI coordinates at the three viewpoints.
- Write at least two items each for numerical differences and visual descriptions.
Self-check
- What does a difference metric mean if the reference image is wrong?
- Which errors does still-image evaluation miss?
Official references
Linked example and execution scope
Example source · Runtime environments and limits
CPU Canvas2D synthetic fixtures and contract checks. This does not provide Spark GPU rendering, training data, or claims about reconstruction quality or GPU performance. The provided example calculates byte-RGB mean squared error (byteRGB_MSE) between 256/128 synthetic splat images from the same camera. The assignment’s mean absolute error and multiple-view comparison require separate implementation.
Source code and records are not overwritten automatically. Resetting an existing example also preserves your notes.
Next learning
Previous: Color Space and Background · Next: Separating Profiling Intervals