---
license: apache-2.0
tags:
  - object-detection
  - wildlife
  - thermal
  - rgb
  - drone
  - ultralytics
  - yolo
datasets:
  - custom
pipeline_tag: object-detection
library_name: ultralytics
---

# Wildlife Detection with YOLOv26 — Drone RGB & Thermal Models

A collection of five **YOLOv26x** models fine-tuned for **wildlife detection in drone imagery**, supporting both **RGB** and **thermal (infrared)** modalities. Models were trained using the [Ultralytics](https://github.com/ultralytics/ultralytics) framework on 1024×1024 px drone images.

---

## Models Overview

| Model | Modality | Dataset | Epochs | mAP50 | mAP50-95 | Notes |
|---|---|---|---|---|---|---|
| `thermal_original` | Thermal | Original thermal dataset | 25 | 0.217 | — | Baseline thermal model |
| `thermal_merged` | Thermal | Original + supplemental thermal data | 23 | 0.390 | 0.250 | Refined thermal model |
| `rgb` | RGB | Full RGB dataset | 72 | 0.946 | 0.655 | Primary RGB model |
| `matched_rgb` | RGB | Matched RGB-thermal pairs | 43 | 0.731 | 0.431 | Cross-modal comparison |
| `matched_thermal` | Thermal | Matched RGB-thermal pairs | 27 | 0.719 | 0.289 | Cross-modal comparison |

> **Matched models** were trained on the same spatially co-registered scene pairs to enable fair modality comparison.

---

## Training Details

All models share the following configuration:

| Parameter | Value |
|---|---|
| Base model | `yolo26x.pt` |
| Image size | 1024 × 1024 px |
| Batch size | 4 |
| Max epochs | 200 |
| Early stopping patience | 20 epochs |
| Optimizer | Auto |
| AMP (mixed precision) | Enabled |
| Close mosaic | Last 10 epochs |
| Data augmentation | RandAugment, erasing (p=0.4), fliplr (p=0.5) |

---

## Repository Structure

```
.
├── README.md
├── inference.py                  # Sample inference code
├── thermal_original/
│   ├── weights/
│   │   ├── best.pt               # Best checkpoint
│   │   └── last.pt               # Last checkpoint
│   ├── args.yaml                 # Training configuration
│   ├── results.csv               # Per-epoch training metrics
│   └── results.png               # Training curves
├── thermal_merged/
│   └── ...
├── rgb/
│   └── ...
├── matched_rgb/
│   └── ...
└── matched_thermal/
    └── ...
```

---

## Quick Start

### Installation

```bash
pip install ultralytics
```

### Load a model and run inference

```python
from ultralytics import YOLO

# Choose your model
model = YOLO("rgb/weights/best.pt")          # RGB drone imagery
# model = YOLO("thermal_merged/weights/best.pt")  # Thermal imagery

# Run inference on an image
results = model("path/to/your/image.jpg")

# Display / save results
results[0].show()
results[0].save("output.jpg")

# Access detections
for box in results[0].boxes:
    print(f"Class: {box.cls.item()}, Conf: {box.conf.item():.2f}, BBox: {box.xyxy[0].tolist()}")
```

### Batch inference

```python
from ultralytics import YOLO
from pathlib import Path

model = YOLO("rgb/weights/best.pt")

# Run on a folder of images
results = model(
    source="path/to/images/",
    imgsz=1024,
    conf=0.25,
    iou=0.45,
    save=True,
    project="detections",
    name="run"
)
```

### Modality comparison (matched dataset)

```python
from ultralytics import YOLO

rgb_model     = YOLO("matched_rgb/weights/best.pt")
thermal_model = YOLO("matched_thermal/weights/best.pt")

# Run both models on co-registered image pairs
rgb_results     = rgb_model("rgb_frame.jpg",     imgsz=1024, conf=0.25)
thermal_results = thermal_model("thermal_frame.png", imgsz=1024, conf=0.25)

print(f"RGB detections:     {len(rgb_results[0].boxes)}")
print(f"Thermal detections: {len(thermal_results[0].boxes)}")
```

See **`inference.py`** for a complete script with CLI argument parsing.

---

## Results Visualizations

Training curves, precision-recall curves, and validation batch predictions are included in each model subdirectory (`.png` / `.jpg` files).

---

## License

Apache 2.0 — see `LICENSE`.