Synthesizing Human Activity for Data Generation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Gathering representative human action, shape, and facial expression data is costly and time-consuming.
• Existing methods like transfer learning and data augmentation are often insufficient for creating diverse datasets.
• The need for robust models necessitates innovative data generation techniques.

Purpose of the Study:

• To propose a semi-automated mechanism for generating and editing visual scenes with synthetic humans.
• To enable the creation of data with greater variability through background modification and 3D avatar adjustments.
• To introduce a two-fold evaluation methodology for assessing the generated data's effectiveness.

Main Methods:

• Development of a semi-automated system for synthetic human data generation and editing.
• Incorporation of features like background modification and manual 3D avatar adjustments.
• Implementation of an evaluation methodology involving an action classifier and segmentation-based mask comparison.

Main Results:

• Generated synthetic avatars demonstrated robustness to occlusion, with recognizable and accurate actions.
• Action classification relied heavily on contextual information beyond pose and movement for precise recognition.
• Challenges were identified in generating avatars for complex activities and achieving clean segmentation masks.

Conclusions:

• The proposed semi-automated mechanism effectively generates diverse synthetic human data for AI training.
• Contextual information is crucial for accurate action recognition in synthetic data.
• Further research is needed to address limitations in complex activity generation and mask precision.