CAnnotator: Photo-Guided Color Annotation for Degraded Ancient Paintings
summary
CAnnotator is an interactive system built over a 2‑year study that links degraded paintings to photo collections, extracts textures with a fine‑tuned tracker, and simulates pigment mixing with a learned model and a novel Color Flow visualization. The pipeline combines Segment Anything for segmentation, a texture‑tracking model fine‑tuned to ~60% precision/recall, and a pigment‑mixing network whose outputs are often indistinguishable from ground truth to non‑expert evaluators. In lab studies CAnnotator reduces color‑distance versus manual annotation and increases user confidence while users spend similar or more time exploring color combinations.
AI-written summary · about
chapters (11)
- 0:00Introduction and motivation
- 1:10Challenges in color annotation
- 2:36Solution overview and workflow
- 3:19Paint annotation view demo
- 4:01Photo reference view demo
- 4:24Pigment mixing view and Color Flow
- 5:08Model evaluation results
- 6:12User study findings
- 7:37Real-world use case
- 8:21Conclusion and future work
- 8:43Q&A: Color accuracy discussion
related talks
08mDynamic Guidance for Decluttering Photographic Compositions
10mRecreating Photography of the 1850s in a Digital World
45mColours on Time: stronger schemes, fewer headaches
22mColour in the Real World
04mLabel Studio
26mAnatomy of Dynamic Color
11mWhere Are My Tomatoes? Extracting Data from Graph Images with AI
31mPicasso, Geometry, Jupyter