DeepForest: Recovering computer vision annotations from historical airborne imagery monitoring

After the 2010 Deepwater Horizon oil spill, airborne surveys produced 18,304 images with 2.6 million birds individually marked using a point-counting tool. The tool baked colored annotation dots directly into screenshot pixels no coordinates were ever saved. I built a six-stage recovery pipeline that extracts dot positions from corrupted screenshots, maps them to original high-resolution photographs via SIFT homography, and exports DeepForest-compatible training annotations. Tested on 34 images spanning 7 years and 10 annotators with zero failures: 70.8% detection accuracy, 98.3% position precision, 3,915 annotations recovered across 21 species. Training experiments confirmed that position accuracy of recovered coordinates not data volume drives downstream detector performance. GSoC deliverables: (1) Scale recovery to all 18,304 images with an optimized feature matching cascade targeting 80%+ geometric precision, (2) a quality-tiered dataset of 250,000–340,000 recovered annotations, (3) a fine-tuned airborne bird detector evaluated per-species on held-out colonies, (4) a tested DeepForest module with documentation, and (5) a blog post documenting the full recovery process including approaches that failed and why.