Data Augmentation-based method among Few-Shot Learning (FSL) approaches
Among the metric-based algorithms that learn similarity between samples and compare the distances between features extracted through samples, MatchingNet, PrototypicalNet, and RelationNet are used as baselines. These methods have lower performance than the many-shot methods, and to improve this, the method of automatically finding the optimal data augmentation strategy from the Data Augmentation Candidate Pool is applied.
Utilization of Augmented Dataset in Classification Phase and Evaluation of Metric-based FSL Model Performance
The first application of automatic increment/decrement technique to FSL algorithm and automation of search mechanism for augmentation-based FSL. EDANet proposal to automatically search for optimal augmentation rules without human intervention. As a result of the experiment, EDANet searches for an optimal data enrichment strategy from the pool of candidate enrichment techniques for metric-based FSL, and improves performance compared to the existing baseline.
EDANet: composed of Augmentation Phase and Classification Phase
Development of learning data construction technology for construction sites
In order to identify workers’ non-compliance with regulations and dangerous/unsafe behaviors with visual intelligence, it is necessary to collect learning data, but collecting data directly from actual construction sites or collecting data through directing is very limited. Therefore, by modeling and directing workers and work sites in a virtual space, it is intended to collect a large number of learning materials without time and space limitations.
Building a game development engine-based virtual environment and creating learning data
By using Unity, the game development engine, a virtual work environment with high similarity to the real work environment is constructed, and by combining data automatic labeling, data augmentation, and learning algorithms, it is possible to determine with high accuracy the non-compliance and risk/unsafe behavior of workers. A deep neural network (DNN) model was trained.
After filming an actual construction site and inserting a 3D Mesh file into Unity, human modeling (animation of walking, standing, etc.) and safety ring modeling allow dangerous/unsafe behavior to be produced in virtual space.
Generate labeling data
Using the Perception Camera function provided by Unity Computer Vision, a machine learning learning tool provided by Unity, it was developed to generate data automatically labeled with people and safety rings (2D Bounding Box and Semantic Segmentation) and perform data labeling and augmentation.
Virtual workspace-based data creation, augmentation and learning process