s1hofmann / piper Goto Github PK

Maybe go for boost instead to lower dependencies?

Each config class could have its own import method to parse config data from a JSON file.

The respective fields would be stored within a separate node given the same name as the respective pipeline step.

If values are missing default values should be used.

Only some steps take an actual mask, where as some take two separate arguments. Therefore "param" would be a more convenient name.

Provide a separate configuration for global pipeline options like output directories for descriptors, labelfiles etc.

The old code relies on a template function to create numeric ranges, this would also be cool to have.

Configuration could be done via a json file or something similar, the file / folder to process as well as the output summary should be passed as parameters.

Currently everything is handeled by a single, large CMakelists.txt file.
Might be better if every module gets its own file.

Better visualization of progress.

Use libProgress

Add support for different machine learning implementations / libraries (e.g. caffe) and the ability to opt them in in the build config

Add possibility to set a working directory for processing.

Slower, but increased accuracy?

Config classes could provide a create method which returns a new correct configured PipelineStep object.

A class to shuffle data for input to the SGD classifier.

Auto generated log messages would be cool, in case you're processing an empty matrix or stuff like that.

Maybe just stick with single channel?
Keep processing the same, just split up on every channel?

Add possibility to store configs as json in order to export hard coded configs.

By now options are bound to one character. Longer options would be beneficial for more parameters.