DIRECTORY
PROJECT BACKGROUND
The global fish consumption has increased by over 100% in the past half decade. Aquaculture currently contributes more than 50% to the global food-use fish production and is expected to continue growing. Rotifer, a crucial starting diet for fish larvae, is critical for fish aquaculture. However, the cultivation of rotifer is labor intensive. Automated rotifer culture is one of the necessary and critical piece of next generation aquaculture.
WATCH THE DEMO!
A demo video of the prototype system is now available!
PROJECT MILESTONES
Aug. 2019 - The project was proposed and initiated!
Nov. 2019 - The first batch of data was collected and annotated!
Dec. 2019 - Project framework and structural pipelines were developed and implemented. Progress include:
- Simple median-based background subtraction for making object proposals
- Object classification based on ensemble of weak classifiers trained on evolutionary constructed features
- IOU tracking and feature matching-based object re-identification.
Jan. 2020 - The project was sponsored by Microsoft Azure AI for Earth Computing Grant! The first deep learning model was implemented and tested.
Feb. 2020 - The project was presented at the Aquaculture America 2020 Conference in Hawaii, HL. [poster]
Coming soon…
- Collect and annotate more data.
- Try all kinds of things to optimize the deep learning model.
- Annotate several videos for evaluating the performance of the rotifer tracking and re-identification models.
- Evaluate and improve the object tracking and re-identification model.
- Develop a model to infer the true class of the tracked objects based on tracking history.
- More ideas later…
PROJECT BLOGs
BLOG#1 - More About Rotifers and A Glimpse of the Rotifer Dataset
The rotifer dataset is consisted of microscopic images and videos of rotifer culture samples, collected and annotated by skillful rotifer culture specialists at the University of Miami Experimental Hatchery. The dataset is prepared for training classic machine learning algorithms as well as the deep learning models to achieve the project objectives.
BLOG#2 - Moving Object Detection
Moving object detection algorithm is able to efficiently generate object proposals. The proposed patches can then be classified by well-trained object classifier into classes of interests.