kailunw/visdrone-yolo26m

🚁 VisDrone YOLO26m Object Detection

YOLO26m model fine-tuned on the VisDrone2019-DET dataset for dense small-object detection in aerial imagery.

---

📌 Overview

This project trains a YOLO26m detector on the VisDrone dataset, which features:

• Dense object distribution
• Small-scale targets
• Heavy occlusion
• Real-world UAV scenarios

---

🧠 Model Details

• Model: YOLO26m (Ultralytics)
• Task: Object Detection
• Input Size: 1024 × 1024
• Classes: 10
• Framework: PyTorch

---

📊 Performance (Validation)

Metric	Value
mAP@0.5	0.5505
mAP@0.5:0.95	0.3407
Precision	0.6453
Recall	0.5431

Best performance achieved at epoch 66.

---

📊 Per-Class AP@50

Class	AP
pedestrian	0.6563
people	0.5164
bicycle	0.3593
car	0.8752
van	0.5726
truck	0.5173
tricycle	0.4503
awning-tricycle	0.2201
bus	0.6956
motor	0.6415

---

⚡ Inference Speed

• 43.3 FPS (~23.1 ms/image)

---

📦 Files

• best.pt → best checkpoint
• last.pt → final checkpoint
• args.yaml → training config
• results.png → training curves

---

🚀 Usage

Install

pip install ultralytics

---

Run inference

from ultralytics import YOLO

model = YOLO("best.pt")
results = model("test.jpg")
results[0].show()

---

Load from Hugging Face

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

model_path = hf_hub_download(
    repo_id="kailunw/visdrone-yolo26m",
    filename="best.pt"
)

model = YOLO(model_path)
results = model("test.jpg")

---

📈 Training Details

• Dataset: VisDrone2019-DET
• Epochs: 200
• Best Epoch: 66
• Image Size: 1024
• Optimizer: MuSGD
• Mixed Precision (AMP): Enabled

---

🧠 Key Insights

• Strong performance on larger objects (car, bus)
• Lower performance on small/rare objects (awning-tricycle)
• Dense scenes make precise localization challenging
• Model achieves real-time inference (~43 FPS)

---

🔧 Future Work

• Increase input resolution (1280+)
• Improve small-object detection
• Balance dataset for rare classes

---

👨‍💻 Author

Kailun Wang
Upenn Robotics