Junior Machine Learning Engineer (Computer Vision)

Built a cascaded YOLOv9 + SAM2 pipeline for zero-shot building/road segmentation in satellite imagery.

What I Did

I built a pipeline that chains YOLOv9 object detection with Meta's SAM2 segmentation model. YOLO's bounding box centroids are used as sparse prompts for SAM2 to produce pixel-accurate masks. I implemented caching strategies to avoid redundant I/O when processing large satellite images.

Impact

The pipeline automated building and road detection for the Humanitarian OpenStreetMap Team, enabling mapping of regions that lacked detailed maps. The caching improvements made the system practical for larger datasets.

What I Learned

I gained experience with object detection (YOLO architecture, anchor boxes, NMS) and promptable segmentation (SAM2's sparse/dense prompting). I learned about geospatial image formats (GeoTIFF, coordinate reference systems) and how to tile large raster images for batch processing. The caching work taught me about serialization formats and I/O optimization.

Key Highlights

• Built a cascaded inference pipeline injecting YOLOv9 bounding-box centroids as sparse prompts into Meta's SAM2 for zero-shot segmentation, implementing efficient data serialization caching strategies to drastically minimize critical I/O latency bottlenecks.