File size: 4,248 Bytes
5dbce6d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 | ---
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`.
|