1 Photo OCR
1.1 Program description and pipeline
1.2 Sliding windows
- take pedestrian detection as an example
using different rectangles to scan the image, step by step-size or stride, say 4 pixels - Text detection
got the result like this
1.3 Character segmentation
1.4 Character classification
2 Getting lots of data: Artificial data synthesis
2.1 Artificial data examples
- use lots of fonts to make synthetic data
- synthesizing data by introducing distortions
- Distortion should be representatioin of the type of noise/distrotions in the test set, eg: background noise, bad cellphone connection sounds
- Usually does not help to add purely random/meaningless noise.
2.2 Discussion on getting more data
- Make sure you have a low bias classifier before getting more data.(Plotting learning curves)
- Always ask “How much work would it be to get 10times as much data as we currently have?”
1)Artificial data synthesis
2)Collect/label it by yourself
3)“Crowd source”(eg: Amazon Mechanical Turk)