harsh-awasthi/bluemockdrone

Drone Frame YOLOv11n Object Detector (v2.0 — Structure-First & Color-Invariant)

Ultra-fast, high-precision YOLOv11 Nano object detector trained to recognize the geometric structure, spoke morphology, circular rim, and central aperture of a 3D-printed drone gimbal / tracking wheel across varied physical environments, lighting extremes, and arbitrary colors.

---

📌 Changes in v2.0 vs v1.0 (Issues of Older Version Solved)

Dimension	Version 1.0 (Previous)	Version 2.0 (Current)
Object Color Bias	Trained almost exclusively on blue objects (Hue jitter $\pm 8$). Suffered severe color-shortcut overfitting.	Color-Invariant: Trained across full 360° hue shifts (red, green, yellow, purple, orange, cyan, blue).
Grayscale / Edge Focus	0% grayscale images. Dependent on color cues.	25% Pure Grayscale: Network is forced to detect pure edges, spokes, and circular hub geometry.
False Positives	Detected arbitrary blue objects (blue denim jeans, shirts, cups, blue walls, mousepads).	Drastically Reduced: Negative background samples (~10% empty labels) and in-scene blue distractors penalize blue false alarms.
Background Realism	Synthetic cutout pasting with visible segmentation artifacts.	Real Scene Continuity: Real camera images augmented with clean constant letterboxing.
Dataset Scale	400 images.	1,000 images (800 train, 200 val) with multi-core CPU augmentation.
Low-Light / Glare	Failed in low-light, shadows, or bright ceiling glares.	Robust detection under severe underexposure, glare, and shadows.

---

🚀 Quick Start & Inference

Python (Ultralytics)

from ultralytics import YOLO

# 1. Load model directly from Hugging Face
model = YOLO("harsh-awasthi/bluemockdrone")

# 2. Run inference on any image, video, or webcam feed
results = model.predict(source=0, show=True, conf=0.45)

ONNX Runtime (Embedded / Raspberry Pi / Edge)

import onnxruntime as ort
import cv2
import numpy as np

session = ort.InferenceSession("best.onnx")
# Ready for high-speed edge deployment on pan-tilt visual servoing systems!

---

🎯 Architecture & Performance

• Base Architecture: YOLOv11n (Nano)
• Input Resolution: 640×640
• Primary Class: drone_frame (Class 0)
• Framework: Ultralytics YOLOv11