Existing datasets for planktonic organisms are numerous, and they are provided from different environment and setup. In this project Sverre is studying the effect and performance of training deep learning architectures over a set of different labeled datasets. He applies methods of transfer learning for plankton classification. 

The deep-learning architecture is trained over multiple planktonic datasets (ex. WHOI, ZooScan and Kaggle) then its classification performance is validated over the SilCam collected data.

Existing planktonic datasets suffer from class imbalance. Trained deep learning architectures tend to favor classification of objects to higher weight classes. This can mislead the segmentation and classification procedure. In this project, Oda is studying recent approaches to solve the class-imbalance and evaluate their performance over datasets of planktonic organisms.

Output classification should have a balanced structure

She presented her work in the Ocean Sciences meeting, February 2020

Oda, Kiese; Saad, Aya; Stahl, Annette. (2020) Towards a Balanced-Labeled-Dataset of Planktons for a Better In-Situ Taxa Identification. Ocean Sciences Meeting 2020 . AGU, ASLO and TOS; San Diego, CA. 2020-02-16 - 2020-02-21.

Sondre is using the Mask R-CNN based on detectron2 framework (developed by Facebook AI research) for object detection and instance segmentation. He created a small dataset for instance segmentation using the VGG annotator tool

He has an accepted paper in Oceans Gulf Coast Conference, October 5-14th, 2020.

Source code available @ https://github.com/AILARON/Segmentation

Project 4: Applying Deep-learning methods of Few-shot Image Recognition to discover new unknown classes

Andreas has developed a new framework that combines few-shot learning with outlier-detection algorithm. The purpose of this framework is to detect in-situ unknown/unseen classes of planktonic images.

Accepted paper in the student competition track of the Oceans Gulf Coast Conference, October 5-14th, 2020

Eivind is applying recent unsupervised deep learning methods to classify planktonic images captured in-situ. Proposed list of classes are then validated by domain expert biologists.

Auto Encoder-Decoder architecture

Accepted paper in the student competition track of the Oceans Gulf Coast Conference, October 5-14th, 2020

Source code available @ https://github.com/AILARON/Unsupervised-classification

Martin is investigating how to apply techniques for active learning, where the network is trained over a small dataset. Results of classification with high confidence are included in the dataset in the training interactions that follow. Approaches of active learning help minimizing the labeling effort of large datasets.

Jonas is investigating recent Deep Learning approaches for instance segmentation to identify planktons and their distribution from time-series images taken in-situ.