Abstract
The project’s goal was to set up a pipeline to easily evaluate the performance of newly generated synthetic data on standard hand pose estimation algorithm. Hence, a choice on datasets and algorithms available publicly was made based on easiness to make them work both independently and together. This involved a need to understand deeply the research-based algorithm in order to modify the code to apply it to the company’s specific needs. Hence, the project involved both software development and research. By the end of the project, the pipeline was fully working and able to demonstrate performance of 3 different datasets on 1 hand pose estimation algorithm. Improvements on the hand pose estimation algorithm were also tested and used.
Challenges
- Understanding research papers and the code linked to it in order to modify algorithms without breaking them.
- Writing conversion scripts to get datasets ready to work as input to different algorithms.
- Trial-and-error new ways to improve algorithmic performance by inventing a new loss function and setting up data augmentation procedures.
Achievements (according to KPIs)
- Pipeline fully working. Set up 3 datasets and 1 algorithm.
- Datagen data reported performance: high pixel mean end point error (EPE) on the STB dataset.
Further development
Continuing to add algorithms to the pipeline.
Add a pleasant GUI.
Find a way to plugin new algorithms with minimum code modification.