vision_dev

Scripts

• split_data.sh

  This script splits a directory full of images into multiple numbered directories with 100 images each

  • Usage

    ./split_data.sh <source directory>
    • source directory: [string] The directory storing the original set of image

Sloth

Sloth is an image tagging tool.

Configuration

Provided, is the sloth configuration file, robosub_config.py, to specialize sloth for Robosub use. Included, are the required annotation conversions for any frameworks used by robosub.

• Annotation types

  • Darknet

    Darknet has its own specific format which requires multiple files, one per image. For this, we have created a special *.darknet format to bridge the gap between sloth's one file system to darknet's multiple file system. The *.darknet format is as follows

    ---
    <class name> : <class number>
    class_name_1 : 0
    class_name_2 : 1
          :
          :
    ---
    >> "filename_1.jpg"
    <class number> <x_position_ratio> <y_position_ratio> <width_ratio> <height_ratio> //Darknet format line
    <<
    >> "filename_2.png"
    <<
        :
        :
    

    • class name: [string] The name of the class within sloth for the tag
    • class number: [integer] The number assotiated with the class name used within darknet
    • x_position_ratio: [0.0...1.0] The ratio between the x position of the center of the rectangle and the image width
    • y_position_ratio: [0.0...1.0] The ratio between the y position of the center of the rectangle and the image height
    • width_ratio: [0.0...1.0] The ratio between the width of the rectangle and the image width
    • height_ratio: [0.0...1.0] The ratio between the height of the rectangle and the image height

    Utilize the sloth_to_darknet.py script to generate the darknet files from the above annotation format.

Scripts

• sloth_to_darknet.py

  Converts from the *.darknet format specified above to multiple files within an output directory and a training list used with Darknet

  • Usage

    ./sloth_to_darknet.py -f <filename> -o <output directory> [-t <training list filename>]
    • filename: [string] The filepath to the *.darknet file from which to extract information.
    • output directory: [string] The directory into which the darknet compatible annotation files are output.
    • (optional) training list filename: [string] The filename of the output training list

• sloth_to_tfrecord.py

  Converts json format files into tenserflow format.

  • Usage

    ./sloth_to_tfrecord.py <config_file> <output directory> --crossvalidation <crossvalidation> --test <testratio> --evaluation <evaluationratio>
    • input [STRING]: resourcename from REST resource or json-file ending in .json
      • .json for offline-usage
    • outputdir [STRING]: folder in which the tfrecord files will be written + labelMap
    • crossvalidation [INT] (default 1): determines amounts of training- & test-datasets that will be generated using the crossvalidation algorithm
      • eg: 10 -> evaluationratio is substracted from all files - the remaining are used to create X sets. Has higher priority than the testratio parameter. crossvalidation == 1 -> data is split into [trainingset / testset / validationset] according to testratio & evaluationratio
    • testratio [FLOAT] (default 0.1): determines the ratio of files will be used as a test record. Will be ignored when crossvalidation parameter is greater than 1
      • eg: 0.1 -> 10% of the labelled files will be written to test.tfrecord
    • evaluationratio [FLOAT] (default 0.2): determines the ratio of files will be used as an evaluation record
      • eg: 0.1 -> 10% of the labelled files will be written to eval.tfrecord