About the OCDS AI Portal
The Office of the Chief Data Steward has created AI powered tools that allow automated processing of images to bring the following advantages
- Scalability: These AI systems can be deployed to annotate massive datasets with minimal additional cost.
- Improved consistency: AI applies the same learned criteria across all images, reducing human variability and bias.
- Continuous improvement: AI can improve over time with more data, leading to progressively better performance.
- Increased speed and efficiency: These systems can process and annotate large volumes of images much faster than manual annotation alone.
- Reduced human labor: AI assistance reduces the manual workload, allowing human annotators to focus on edge cases and quality control.
Technical Details
This system is based on a Detectron model with a ResNet-50 backbone to identify and count commonly harvested fish species in imagery collected from electronic monitoring equipement on commercial fishing vessels.
The training data was provided by collected over 9 cruises and consists of 477,889 center points (fish) in 3,362 images annotated by trained staff. The model was trained on an 70/15/15 train/validation/test split at the image level.
This is a RetinaNet model fine-tuned from the Detectron2 object detection platform's ResNet backbone to identify 21 benthic morphotaxonomic categories drawn from underwater remotely operated vehicles.
The data is collected from FathomNet and DFO internal data and consists of 10,377 images that contain a total of 31,793 localizations. The model was trained on an 70/15/15 train/validation/test split at the image level.
This is a Detectron model with a ResNet-50 backbone that estimates the number of fish in underwater shoals and schools in imagery collected from camera equipement during underwater surveys.
The data was sourced from the publicly available IOCfish5K dataset which consists of 5,637 high-resolution images containing a total of 1,659,024 annotated center points (fish). The model was trained on an 70/15/15 train/validation/test split at the image level.
This is a YOLOv6 model fine-tuned from a ResNet-50 backbone to identify Ghost Gear (Abandonned Lobster Traps) from collected Side-scan sonar imagery.
The finetuning data was provided by CSR GeoSurveys Ltd. and consists of 1,009 images that contain a total of 1,568 localizations of lobster traps, 1,068 localizations of ropes and 66 localizations of non-gear (negatives). The model was trained on an 70/15/15 train/validation/test split at the image level.