Active Labeling Project

sellProject sellDFKI sellAugmented Vision sellDeep Learning sellActive Learning sellData Annotation

Active Learning for Optimal Data Labeling (Project)

Nikita Khutorni, Muhammad Saif Ullah Khan
Paper

Abstract

In this project, we developed active learning capabilities for the Computer Vision Annotation Tool (CVAT) to speed up keypoint data annotation. Using the COCO dataset, we designed a pipeline to add new spine keypoint annotations to the dataset. This was achieved through our web-based Continuous Active Learning System (CALS), which works with CVAT to enable efficient data annotation. This significantly reduces the manual labor involved in data annotation by enabling efficient identification and annotation of the most informative instances. Although a comprehensive evaluation of CALS’ usability was beyond the scope of this study due to time constraints, preliminary accomplishments indicate that CALS successfully simplifies AL workflow interactions for non-technical users. The system facilitates data upload and annotation through an accessible web interface, meeting its primary goal of enhancing user engagement in HPE dataset development. Future work will extend CALS’ application beyond HPE, improve query strategy selection, and refine the system’s file system abstraction for better scalability. This contribution is poised to streamline the process of enriching HPE datasets with critical anatomical keypoints, thereby supporting advanced applications in medical analysis, sports science, and interactive technology.

Topic

In this project, we will extend an existing dataset with new annotations using active learning. This involves training a neural network to predict the new annotations we want to add in the dataset, manually refining them in batches, and re-training the network. This network tells us which data samples to review in the next batch. This is an iterative process which is repeated until the whole dataset is done.

Tasks

  1. Create an initial dataset using pretrained models.
  2. Implement an active learning pipeline to aid in annotation.
  3. Manually review/refine predictions, finetune network, predict, repeat. This step involves significant amount of manual data annotation.