Problem Description & Pipeline
Photo Optical Character Recognition
- Text detection
- Character segmentation
- Character classification
- （spelling correction)
Getting Lots of Data and Artificial Data
- Use different fonts combined with random background to generate large dataset.
- Introduce artifical distortions.
- Make sure we have a low bias classifier before expending the effort. (keep increasing the number of features/number of hidden units in neural network)
- How much work would it be to get 10x as much data?
- artificial data synthesis
- collect/label it yourself
- crowd source (Amazon Mechanical Turk)
Ceiling Analysis: What Part of the Pipeline to Work on Next
Good tool to find worthy step to work on. Check the final accuracy step by step by assuming 100% accurate performance for steps.