Text spotter for different map styles

[zekun-li]

• 1. Extract patch-level feature with trained contrastive model for >50K map images
• 2. Compute map-level features from patch-level features (either clustering or simple averaging)
• 3. Divide (cluster) map-level features into <10 groups
• 4. Train cycleGAN model to learn style transfer with OSM and rumsey maps as input (# models depends on # clusters)
• 5. Produce synthetic maps backgrounds with cycleGAN with OSM maps as input (synthetic map backgrounds do NOT contain text)
• 6. Generate text layer with QGIS using various font sizes, styles , orientation and curving
• 7. Combine synthetic map background and text layer to form complete map
• 8. Calculate the ground-truth bounding polygon & text string for text layer ,and save output in TESTR compatible format (some effort required to handle truncated text)
• 9. Fine-tune TESTR with the synthetic map data
• 10. Evaluate the results and compare with other TESTR versions