quailrobot/drone-synth-v2-sample

QM Synthetic Drone Detection v2 (with bird decoys) - Free Sample

<!-- qm-buy:start --> Buy the full commercial edition: $29 USD -> [Polar checkout, instant download](https://buy.polar.sh/polar_cl_7GKayFgG8OxqQu58QUexVwOM0GgEI9krJCN0Q39jc0n) Also on Gumroad. This free sample is non-commercial (CC BY-NC-SA 4.0). The paid full edition has a commercial licence.

Custom dataset of YOUR object ($249)

Need data of YOUR object? Custom synthetic dataset, $249 USD -> [order on Polar](https://buy.polar.sh/polar_cl_AQu6LzRtqiQmKPlgPt0zePBtvWjJI4vULeFMQ4NLHH9)

• —What you get: 2,000 labelled photoreal synthetic images (640x640 JPEG) of your own object or scenario (product, part, tool, drone, package, defect...), up to 3 classes, YOLO bounding boxes + data.yaml, train/val/test split, quality report
• —Licence: commercial use allowed
• —Price: $249 USD one-time; one round of adjustments included
• —Delivery: typically 3-5 business days after we receive your reference photos + rough dimensions
• —Optional sim-to-real test: send ~200 of your own labelled real images and we report how much the synthetic data improves a detector on them
• —Refund: full refund if we cannot deliver your request (14-day refund policy)
• —Limits: only objects you own or are allowed to use; no weapons or anything meant to harm people; no copied third-party 3D assets
• —Order URL: https://buy.polar.sh/polarclAQu6LzRtqiQmKPlgPt0zePBtvWjJI4vULeFMQ4NLHH9
• —Details + contact: https://quailrobot-quailmodel.static.hf.space

Summary

Synthetic sky images of quadcopter drones with YOLO bounding boxes and unlabelled bird decoys, trees and buildings, for training drone detectors; the paid full edition can be used commercially.

• —Best for: pre-training single-class drone detectors, small-object detection research
• —Not for: deploying a detector trained on synthetic images only

Quick start

# pip install huggingface_hub ultralytics
from huggingface_hub import snapshot_download
path = snapshot_download("quailrobot/drone-synth-v2-sample", repo_type="dataset")
from ultralytics import YOLO
YOLO("yolov8n.pt").train(data=f"{path}/data.yaml", epochs=50, imgsz=640)

Facts

• —Task: object detection (YOLO format), single class
• —Labels: 1 class: 0 drone; birds are unlabelled hard negatives; empty label file = true negative
• —Full edition size: 5,000 images 640x640 (train 3,976 / val 508 / test 516) with YOLO labels (zip 93.6 MB)
• —Free sample size: 300 images (train 235 / val 31 / test 34)
• —Format: 640x640 JPEG images + YOLO txt labels + data.yaml (Ultralytics-ready)
• —What's included (full edition): images/{train,val,test}, labels/{train,val,test}, data.yaml, README.md, LICENSE.txt
• —Price: $29 USD; checkout may display the equivalent in your local currency
• —Buy URL (primary): https://buy.polar.sh/polarcl7GKayFgG8OxqQu58QUexVwOM0GgEI9krJCN0Q39jc0n
• —Buy URL (secondary, also on Gumroad): https://quailcraft1.gumroad.com/l/synthetic-drone-detection-v2
• —Licence (full edition): LicenseRef-QuailModel-Commercial (commercial use allowed, no resale of the data)
• —Licence (free sample): CC-BY-NC-SA-4.0
• —Validation: Sim-to-real test (YOLOv8n, 1,000 held-out real drone photos from the Seraphim dataset, CC BY 4.0, evaluation only): 200 real images alone = mAP50 0.719 / mAP50-95 0.367 (mean of 3 seeds; range 0.678-0.744 / 0.337-0.384). Pre-training on QuailModel synthetic (Drone v2 + Airspace v3) then fine-tuning on the same 200 real images = mAP50 0.747 / mAP50-95 0.395 (1 seed): the mAP50 gain is within seed noise; mAP50-95 is +2.8 points. Our newer photoreal Drone v5 gives a consistent +4.0 points mAP50-95 (+11%) across 3 seeds and lower variance. 2,000 real images: 0.807 / 0.491. Synthetic data does not replace real data - use it to pre-train.
• —Data source: 100% synthetic, generated by QuailModel with AI assistance (generator code written with an AI model)
• —Catalog (all QuailModel datasets, catalog.json, llms.txt): https://quailrobot-quailmodel.static.hf.space
• —Last updated: 2026-10-10

Validation

Sim-to-real test (YOLOv8n, 1,000 held-out real drone photos from the Seraphim dataset, CC BY 4.0, evaluation only): 200 real images alone = mAP50 0.719 / mAP50-95 0.367 (mean of 3 seeds; range 0.678-0.744 / 0.337-0.384). Pre-training on QuailModel synthetic (Drone v2 + Airspace v3) then fine-tuning on the same 200 real images = mAP50 0.747 / mAP50-95 0.395 (1 seed): the mAP50 gain is within seed noise; mAP50-95 is +2.8 points. Our newer photoreal Drone v5 gives a consistent +4.0 points mAP50-95 (+11%) across 3 seeds and lower variance. 2,000 real images: 0.807 / 0.491. Synthetic data does not replace real data - use it to pre-train.

Price & licence

• —Full edition: $29 USD. One-time payment, instant download after checkout: https://buy.polar.sh/polarcl7GKayFgG8OxqQu58QUexVwOM0GgEI9krJCN0Q39jc0n (also on Gumroad: https://quailcraft1.gumroad.com/l/synthetic-drone-detection-v2)
• —QuailModel Commercial Dataset Licence (SPDX: LicenseRef-QuailModel-Commercial): you may train, evaluate and ship models, including in commercial products. You may not resell or redistribute the dataset itself.
• —Free sample (this page): CC BY-NC-SA 4.0 - free for non-commercial use.

Limitations

• —Synthetic alone does not replace real data (mAP50 0.260 synthetic-only): pre-train, then fine-tune on real images.
• —Low-poly clutter, no motion blur, quadcopters only (no fixed-wing class - see Airspace v3). <!-- qm-buy:end -->

Synthetic sky images of quadcopter drones with YOLO bounding boxes and unlabelled bird decoys, trees and buildings, for training drone detectors; the paid full edition can be used commercially.

What's inside

• —This free sample: 300 images (train 235 / val 31 / test 34)
• —Full commercial edition: 5,000 images 640x640 (train 3,976 / val 508 / test 516) with YOLO labels (zip 93.6 MB)
• —Task: object detection (YOLO format), single class
• —Labels: 1 class: 0 drone; birds are unlabelled hard negatives; empty label file = true negative
• —Format: 640x640 JPEG images + YOLO txt labels + data.yaml (Ultralytics-ready)

Contents

• —Image size: 640x640 JPEG. Labels: YOLO txt, one class (0 = drone). Images with an empty label file are true negatives (50 of 300).
• —data.yaml included - train directly with Ultralytics YOLO.
• —Box size distribution (fraction of image width): median 0.074, 20% of boxes are smaller than 16 px.

How it was made

Original 3D models, procedurally generated and rendered with a physically based renderer under real-world lighting, with realistic camera effects. Labels are computed exactly from the 3D scene (no hand labelling).

Validation

Sim-to-real test (YOLOv8n, 1,000 held-out real drone photos from the Seraphim dataset, CC BY 4.0, evaluation only): 200 real images alone = mAP50 0.719 / mAP50-95 0.367 (mean of 3 seeds; range 0.678-0.744 / 0.337-0.384). Pre-training on QuailModel synthetic (Drone v2 + Airspace v3) then fine-tuning on the same 200 real images = mAP50 0.747 / mAP50-95 0.395 (1 seed): the mAP50 gain is within seed noise; mAP50-95 is +2.8 points. Our newer photoreal Drone v5 gives a consistent +4.0 points mAP50-95 (+11%) across 3 seeds and lower variance. 2,000 real images: 0.807 / 0.491. Synthetic data does not replace real data - use it to pre-train.

Limitations

• —Synthetic alone does not replace real data (mAP50 0.260 synthetic-only): pre-train, then fine-tune on real images.
• —Low-poly clutter, no motion blur, quadcopters only (no fixed-wing class - see Airspace v3).

Best for / not for

Best for: pre-training single-class drone detectors, small-object detection research. Not for: deploying a detector trained on synthetic images only.

Credits

Lighting environments: Poly Haven HDRIs (CC0), credited.

Licence

Sample edition: CC BY-NC-SA 4.0 (non-commercial). The full commercial edition is sold by QuailModel (see the buy link).

Disclosure

Generated by QuailModel with AI assistance (generator code written with an AI model); all data is synthetic / computer-generated. Validate on your own real data before production use.