yolov4-grouse

Model Card for Sage-Grouse Detection Model (YOLOv4)

This model detects sage-grouse in infrared drone imagery using YOLOv4.

Model Details

Model Description

This model is a YOLOv4 object detection model trained to identify sage-grouse in infrared drone videos and images. It is designed to aid in wildlife monitoring and conservation efforts by automating the detection process.

• Developed by: Ilya Buzytsky / Bias Intelligence Inc.
• Funded by: Bias Intelligence Inc.
• Shared by: Ilya Buzytsky / Bias Intelligence Inc.
• Model type: Object Detection (YOLOv4)
• Language(s) (NLP): N/A (Computer Vision)
• License: GPL-3.0

Model Sources

• Repository: https://github.com/biasintelligence/INDECS
• Paper: N/A (in peer review)

Uses

Intended Use

• Wildlife Monitoring: Automating sage-grouse detection for conservation efforts.
• Environmental Research: Studying habitat conditions using infrared imaging.
• Drone Surveys: Assisting biologists and researchers in identifying populations.

Out-of-Scope Use

• The model is not suitable for real-time processing in low-latency applications.
• The model is not designed for general-purpose object detection or use on non-infrared imagery.
• The model may not generalize well to infrared cameras or environments different from those in the training data.

Bias, Risks, and Limitations

The model’s performance may vary depending on the quality and characteristics of the infrared imagery. Factors such as image resolution, lighting conditions, and occlusions can affect detection accuracy. Additionally, because the model was trained on a specific dataset, it may not generalize well to significantly different geographies, environments, or sensor types.

Recommendations

Users should be aware of these limitations and carefully evaluate the model's performance on their specific dataset. It is recommended to:

• Use high-quality infrared imagery.
• Validate results with manual inspection.
• Fine-tune on additional data if necessary.

How to Get Started with the Model

Follow the installation and usage instructions provided in the INDECS repository.
Run the process_images.py script to process infrared images for sage-grouse detection.

Training Details

Training Data

• Dataset Size: 450 infrared drone images.
• Source: Extracted from 12 drone flight videos captured across 11 distinct locations.
• Flight Patterns: Combination of linear and POI overflight patterns at 150-200 feet altitude.
• Resolution: Images were processed at 640×512 px (native video resolution).
• Annotation: Bounding boxes were manually labeled for each sage-grouse instance.

Training Procedure

Preprocessing

• Frame Extraction: Images were extracted from videos at their original resolution.
• Data Augmentation: None.

Evaluation

Metrics

The model was evaluated using Mean Average Precision (mAP) as the primary metric.

Results

• [email protected]: 98.42%
• Precision: 98%
• Recall: 98%
• F1-Score: 98%
• Average IoU: 81.85%

Summary

The model achieved an mAP of 98.42% on the test set, demonstrating high detection accuracy with strong localization (81.85% IoU) for sage-grouse in infrared aerial images.

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Used: NVIDIA RTX 3090 GPU, 16-core AMD Threadripper, 128 GB RAM.
• Training Duration: ~10 hours.
• Training Environment: Local workstation.

Technical Specifications

Model Architecture and Objective

• Architecture: YOLOv4 object detection network.
• Loss Function: CIoU loss.
• Non-Maximum Suppression (NMS): GreedyNMS (β = 0.6).

Compute Infrastructure

Hardware

• GPU: NVIDIA RTX 3090
• CPU: 16-core AMD Threadripper
• RAM: 128 GB

Software

• CUDA: 11.8
• cuDNN: 11.4
• Darknet Framework
• Python Version: 3.8