TL;DR
This paper tackles the critical issue of seabird bycatch in New Zealand fisheries by developing a deep learning model to automatically detect seabirds in real-world fishing footage. Their fine-tuned YOLOv8 model achieves impressive accuracy (mAP@50 of 0.9926), paving the way for automated monitoring and conservation efforts.
What Problem Does This Paper Address?
New Zealand is a global hotspot for seabirds, but commercial fishing operations unfortunately lead to accidental bycatch, harming these populations and impacting their long-term sustainability. Manually reviewing vast amounts of video footage from fishing vessels to identify these interactions is impractical and time-consuming. This paper aims to automate this process by creating a robust object detection system capable of identifying seabirds in the challenging, unconstrained environment of active fishing operations.
Methods
The researchers employed a deep learning approach using the YOLO (You Only Look Once) object detection framework, specifically focusing on YOLOv8 and YOLOv9 architectures. Their methodology involved:
- Dataset Creation: A new dataset was curated from video footage provided by Fisheries New Zealand, featuring raw still images from onboard cameras. This dataset was manually annotated with bounding boxes around seabirds.
- Transfer Learning: Pre-trained YOLO models (YOLOv8 and YOLOv9) were fine-tuned on the newly created seabird dataset. This process adapts the models, originally trained on general object detection (COCO dataset), to the specific task of seabird identification.
- Hyperparameter Tuning: A grid search method was used to systematically tune key hyperparameters (learning rate and momentum) for the top-performing YOLO models to optimize their performance.
- Model Selection: The best-performing YOLOv8 and YOLOv9 models were identified and compared.
Datasets
- FNZ Dataset: A newly created dataset comprising 8402 annotated images extracted from video footage captured by onboard cameras on commercial fishing vessels. This dataset is unique as it represents real-world marine interactions, unlike many existing datasets focused on controlled environments.
- COCO 2017 Dataset: Used for pre-training the YOLOv8 and YOLOv9 models. This dataset contains images from 80 different object classes.
Key Insights and Ideas
- Real-world Marine Scenarios are Crucial: The paper highlights that models trained on controlled datasets struggle in the dynamic and unconstrained environment of fishing vessels. Their FNZ dataset, capturing these real-world conditions, is key to developing a practical solution.
- YOLOv8's Efficiency: The YOLOv8 architecture, particularly the "large" variant, demonstrated a strong balance between performance and computational efficiency, making it a suitable choice for potential real-time applications.
- Fine-tuning is Essential: Simply using pre-trained models on the FNZ dataset was insufficient. Transfer learning and subsequent hyperparameter tuning were critical to achieving high accuracy.
- mAP@50-95 as a Robust Metric: The paper emphasizes the importance of the mAP@50-95 metric, which considers a range of Intersection over Union (IoU) thresholds, for evaluating performance in more challenging detection scenarios.
Key Results
The fine-tuned YOLOv8 large model achieved state-of-the-art performance on the FNZ test set:
- mAP@50: 0.9926 (exceeding their 90% target)
- mAP@50-95: 0.9147
- The model demonstrated effectiveness in detecting seabirds at varying distances and in different lighting conditions.
- A basic interactive dashboard was developed to visualize detection results, serving as a proof-of-concept for real-world application.
Limitations and Future Work
- Model Uncertainty: The current model struggles with detecting smaller seabirds further away from the camera, especially under poor lighting conditions.
- Dataset Diversity: The FNZ dataset lacks diversity in terms of time of day, weather conditions, and flock sizes, which limits the model's robustness.
- Single-Class Detection: The model is currently limited to detecting a single class ("Seabird") and cannot distinguish between different species.
Future work suggested includes:
- Modifying the YOLO architecture to better detect smaller seabirds and improve performance under various weather and lighting conditions.
- Expanding the FNZ dataset to include more diverse conditions and different seabird species.
- Developing a multi-species seabird detection model.
- Integrating object tracking algorithms (like ByteTrack or DeepSORT) to track and count individual seabirds in video clips.
Summary updated