YOLOv8 (COCO) on drone/aerial photos finds only cars: COCO's 80 classes have no swimming pool, tree, roof or solar panel; use open-vocabulary YOLO-World or OWLv2
Fix: for aerial property features use an open-vocabulary detector with text prompts. With Ultralytics YOLO-World: m = YOLOWorld('yolov8s-world.pt'); m.set_classes(['swimming pool', 'tree', 'solar panel', 'car']); m.predict(img). Or with Hugging Face OWLv2: Owlv2Processor/Owlv2ForObjectDetection.from_pretrained('google/owlv2-base-patch16-ensemble') with text=[['a swimming pool', 'a tree', 'a solar panel']].
What we saw: stock yolov8n.pt on 5 HackPrinceton Matterport drone photos (oblique DJI shots at ~38 m, gimbal pitch -31°) detected 23 vehicles and nothing else. That's expected. The checkpoint's model.names has 80 COCO classes; the relevant ones are 2: car, 7: truck, 8: boat, 25: umbrella, 58: potted plant (COCO category id 64). There's no pool, tree, roof, building or solar panel class. Filtering detections for 'potted plant' as a stand-in for trees gives near-zero hits outdoors, and that's no reason to lower conf.
Practical notes:
- Open-vocabulary models are much heavier on CPU than yolov8n. Downscale (e.g. longest side 1280) or tile the 5280x3956 drone frames, and expect seconds per image.
- Prompt wording matters ('swimming pool' vs 'pool'). Sanity-check by hand on a few frames where you know there is a pool before trusting counts.
- COCO models are trained mostly on ground-level views. Even cars from 40-60 m oblique are small, so imgsz=1280 or tiling helps.
Verify: print(YOLO('yolov8n.pt').names) before designing a pipeline around a class you assume exists.