LibreYOLO9P2s-visdrone

LibreYOLO9P2s-visdrone: small-object aerial detector (research preview)

yolo9_p2 is a LibreYOLO model family: YOLOv9 with an added stride-4 (P2) detection scale. Stock YOLOv9 detects at strides 8/16/32; the P2 head adds a fourth scale at stride 4, so objects in the 4-16 px range land on a grid fine enough to be found. This checkpoint is the s size (7.2M params) trained on VisDrone2019-DET at 768 px.

Research preview. These weights are trained on VisDrone2019-DET, which is licensed CC BY-NC-SA 3.0, so they are for non-commercial use only and are not covered by LibreYOLO's permissive license. The checkpoint detects the 10 VisDrone aerial classes, not COCO.

Results (VisDrone2019-DET val, 548 images, pycocotools)

In a controlled A/B (same recipe, resolution, and transfer init; the only difference was the P2 head), yolo9_p2-t beat stock yolo9-t by +49% AP_small, with the textbook size signature: large wins on small objects, slight loss on large ones. Across the project (P2 head + higher resolution + bigger size), small-object AP roughly doubled (0.070 → 0.141).

The full per-epoch metric history is in results.csv.

Honest scope notes

• Match the architecture to the arena. On COCO, the same architecture is worse than stock YOLOv9 (COCO "small" is 16-32 px, already covered by the stride-8 level). Use this for aerial/tiny-object imagery where objects fall below ~16 px; it is not a general-purpose detector.
• Not SOTA. Transformer-based small-object detectors (e.g. TinyFormer) reach higher AP on VisDrone at lower resolution. The value here is a simple, fast CNN that substantially improves its own small-object AP.
• Single seed. Every number is one training run; treat ±1 point as noise.
• Evaluate at 768. Train/test resolution mismatch tanks the numbers.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreYOLO9P2s-visdrone.pt")  # auto-downloads from this repo
results = model.predict("aerial.jpg", imgsz=768, conf=0.25)
results[0].show()

The yolo9_p2 family is merged on dev but not yet in a PyPI release. Until it is, install LibreYOLO from source: pip install git+https://github.com/LibreYOLO/libreyolo@dev.

Classes: pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, motor.

Reproduce

1. Download VisDrone2019-DET (train + val) from the official source (https://github.com/VisDrone/VisDrone-Dataset). The images are not redistributed here.
2. Convert to YOLO layout with build_visdrone.py (clean-room converter written from the public annotation spec).
3. Train with train_visdrone.py, the exact recipe that produced this checkpoint (60 epochs, transfer init from stock LibreYOLO9s.pt, lr0=0.005, mosaic/mixup off, hsv+flip on, max_labels=600, imgsz=768, effective batch 16). Roughly 10 min/epoch on an RTX 5070 Ti.

Training notes that matter: the family default lr0=0.01 diverges on transfer init; use 0.005. Mosaic hurts on VisDrone (tiling shrinks tiny objects below detectability); keep it off.

Attribution

• Dataset: VisDrone2019-DET, AISKYEYE team, Lab of Machine Learning and Data Mining, Tianjin University, China (CC BY-NC-SA 3.0).
• Base architecture: YOLOv9 (MIT, MultimediaTechLab/YOLO); P2 extension by LibreYOLO.
• Transfer init: stock LibreYOLO9s.pt (converted upstream YOLOv9-s weights).