Skip to content

Latest commit

 

History

History
38 lines (25 loc) · 2.48 KB

prediction.md

File metadata and controls

38 lines (25 loc) · 2.48 KB
Raw

Prediction

We plan to use two levels of detections:

1- one to predict the location of the scenses in the mansucript pages 2- one to predict location and classification of the descriptors in each of the scenes

Prediction of location of Mandragore Scenes

We expect to use dh_segment, with post processing Ornaments

Tools vailable for Ornament predictions

  • root folder for programs: exps/Ornaments
  • programs:

    • ornaments_data_set_generator.py -i image-folder -o data-folder : generate a classes.txt for the training/prediction with one class and one color (0,255,0). And copy files from image-folder + labels into data-folder, spreading into 3 differents sets (train, test or validation). Input-folder is pointing the images folder. However it expects that the directory is sibling to a 'labels' directory that contains labels. I mean a 2 color png image. => check DEMO to see how to generate these labels from known localization of scense in (x,y,w,h) expected to by used with the train.py program (or for evaluation score after post-processing ?)

    • ornaments_process_sets.py -m <model-dir> -i <input-data-folder> -o <output-prediction-dir> -pp [flag for post-process-only]

      • <model-dir> should be either loc of sownloaded model, or nodel trained thanks to TRAIN.py on the set generated above

      • <input-data-folder> should be the set of images we want the detection for. if not option -pp the program will launch a prediction on the set of images, before running the post-process for ornaments

      • <output-prediction-dir> the output for predictions (will contain at least XML docs of the predictions)

      • Some option allows to change the parameters of post-processing only : --post_process_params <json-file> wher json file contains a section params with the values for post-precessing.

        • Default values are: {"threshold": -1, "ksize_open": [5, 5], "ksize_close": [5, 5]}

    • ornaments_process_eval.py -gt <input-data-folder> -d <npy-prediction-dir> -o <output-eval-dir> -p <param-file>. Compute the scores of prediction against each of the params of the param file (default is provided if param file is not)

dh_segment models and predictions

TO BE DESCRIBED

  • reenforcement on existing models (vgg16 and resnet50)
  • use post processing to extract segments from the prediction on the image
  • several cases : lines, polylines, boxes
  • ornaments is using a box detection - I guess on the "white parts of the image, that is not outside - not the background)