🐨 Koala Detection in Thermal Imagery using RT-DETR (ONNX)
🔍 Model Overview
This model performs koala detection in thermal imagery using a fine-tuned RT-DETR architecture, exported to ONNX for efficient inference.
Unlike standard object detection systems trained on RGB images, this model is specifically designed to operate on thermal camera data, enabling reliable detection in challenging real-world environments such as:
- Night-time conditions
- Dense vegetation
- Low-visibility scenarios
Thermal imaging captures heat signatures rather than visible light. This makes it particularly effective for wildlife detection because:
- Animals appear as distinct hot regions against cooler backgrounds
- Detection works even in complete darkness
- Reduced sensitivity to color, shadows, and lighting variations
This is especially useful for detecting koalas in large forestry environments where RGB-based systems often fail.
🧠 Model Details
- Architecture: RT-DETR
- Format: ONNX
- Input size:
640 × 640 - Output shape:
[1, 300, 5]
⚙️ Inference Details
Preprocessing
- Image resized to 640 × 640
- RGB conversion
- Normalization to [0, 1]
- No letterboxing (direct resize)
Postprocessing
- Convert normalized cx, cy, w, h → pixel coordinates
- Apply confidence threshold
- Draw circle-based visualization
🎯 Applications
- Forestry monitoring and compliance
- Wildlife conservation
- Autonomous ground robots
- Drone-based thermal surveys
- Night-time animal detection
⚠️ Limitations
- Trained on thermal imagery only
- Single-class model (koala only)
- May produce overlapping detections
- Performance depends on thermal camera quality
Training Details
- Dataset: Custom thermal image dataset of koalas
- Task: Object detection
- Model: RT-DETR
- Export: PyTorch -> ONNX
Real time detection
 |
 |
 |
 |
Acknowledgements
- RT-DETR implementation
- ONNX Runtime for efficient inference
Contact
If you are working on similar problems in:
- wildlife monitoring
- robotics
- thermal vision systems
fell free to contact