Sample Data Request: Recovering Bird Annotations from Deepwater Horizon Imagery

[vickysharma-prog]

Hi Ben, Henry, Josh, and Ethan,
I'm Vicky, a Computer Science and Mathematics student interested in Proposal 1: Recovering computer vision annotations from historical airborne imagery for GSoC 2026.

Why This Project?

This project aligns closely with my interests in computer vision and building data pipelines for machine learning. I've been working with object detection and segmentation models (including the Segment Anything family) and exploring how they can be used to identify objects in complex scenes such as aerial imagery.
Through my DeepForest contributions, I've also become familiar with the repository's training pipeline, RetinaNet backbone, and the CSV annotation format used for training. Recovering the dot annotations from the Deepwater Horizon surveys into bounding boxes feels like a natural combination of these skills.
I'm particularly interested in building a pipeline that converts historical dot annotations into ML-ready datasets, enabling this valuable ecological data to be reused for modern computer vision models.

My DeepForest Contributions

• Add regression test for HuggingFace model label_dict loading #1307
• Fix broken github-script action hash in automate-waiting-labels workflow #1340
• Add HuggingFace token warning and authentication docs #1323
• Expose detections_per_img and topk_candidates in config #1314
• Fix low recall when limit_val_batches is set #1298
• Add regression test for read_file GeoDataFrame return type (#1186) #1339
  These contributions helped me better understand the annotation pipeline and training workflow in DeepForest.

Prototype Progress

I've started exploring a hybrid pipeline for recovering bird annotations using both available data sources.

**Track A – Dot Recovery (from screenshots)**
Screenshot with dot annotations
↓
Detect dot positions (color / shape detection)
↓
Extract coordinates
↓
Bird count (ground truth)
**Track B – Bird Detection (from original imagery)**
Original aerial image
↓
Segmentation / detection model 
↓
Bounding boxes + masks
**Alignment & Export**
Dot coordinates + detected boxes
↓
Match dots ↔ detections
↓
Validate count consistency
↓
DeepForest-compatible CSV
↓
Training dataset
This combines **dot annotations (location)** with **model detections (object extent)** to produce validated, ML-ready annotations.
Early experiments on both tracks are promising.

Sample Data Request

To take the prototype further, it would be very helpful to have a few example images from the Deepwater Horizon dataset. Even 5–10 samples would be extremely useful:

• Original aerial image
• Screenshot with dot annotations
• Bird count metadata (if available)

Question

I had one question about the intended direction. Should the primary focus be:

1. Recovering bird positions directly from the dot annotations in the screenshots, or
2. Detecting birds from the original imagery and aligning them with the dot counts?

I want to make sure the prototype is moving in the right direction before going deeper.

Contact

• GitHub: vickysharma-prog
• LinkedIn: https://www.linkedin.com/in/vickysharmasynap/
• Email: raghunathsharma296@gmail.com
• Discord: vickysharma6432

Thanks for your time!
Vicky