FilippTrigub/yolov11x-drone-finetuned

YOLOv11x Drone Detector — Fine-tuned

YOLOv11x fine-tuned for drone detection on a combined dataset of 17,351 images from two public Roboflow datasets. Extends `doguilmak/Drone-Detection-YOLOv11x` with broader training data for improved generalisation on diverse backgrounds and ground-camera perspectives.

Usage

from ultralytics import YOLO

model = YOLO("weights/best.pt")
results = model("your_image.jpg", conf=0.25)
results[0].show()

For ONNX inference:

import onnxruntime as ort
import numpy as np

# Input: [1, 3, 640, 640] float32 RGB, letterboxed
# Output: [1, 300, 6] — NMS baked in, format [x1, y1, x2, y2, score, class]
session = ort.InferenceSession("weights/best.onnx")

Model details

Training

Training data — combined from two Roboflow datasets:

All labels were unified to a single drone class (index 0).

Evaluation

Evaluated on the combined validation set (2,320 images, conf=0.25, IoU=0.5):

The base model's higher numbers are on a smaller, in-distribution validation set. This model was evaluated on a harder, more diverse out-of-distribution set — a better proxy for real-world performance across varied backgrounds.

Intended use

Real-time drone detection from fixed or mobile cameras. Designed for scenarios with varied backgrounds (sky, treeline, urban). Not validated for maritime or night-vision use.