This page evaluates cross-modal retrieval performance across different lighting modalities. For each query sample in one modality (e.g., RGB image), the system retrieves the most similar samples from a target modality (e.g., environment map) based on learned embeddings. The visualization shows both top-k best matches and worst matches for qualitative analysis.
| Component | Description | Visual Indicator |
|---|---|---|
| Modalities Row | Display of all available lighting modalities for each sample (RGB, environment map, text, etc.) | Blue border | Shows content for each modality |
| SH Visualizations | Spherical Harmonics representations: predicted SH, ground truth SH | Green sections | Shown when envmap_dir provided |
| Query Modality | The source modality used to search (e.g., RGB, text, environment map) | Red border | Label shows modality type |
| Top-K Matches | Most similar retrieved samples ranked by cosine similarity score | Rank 1 Rank 2 Rank 3+ |
| Worst Matches | Least similar retrieved samples for failure case analysis | Worst Rank | Shown when enabled |
| Similarity Score | Cosine similarity between query and retrieved embeddings (higher = more similar) | Displayed below each retrieved item (range: -1 to 1) |











































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































