README.md

Hermes Bib Detection

This script detects bibs for images in a given directory. These images can either be original raw images or cropped images using the person detection script.

Installing

Use the setup.sh script to install.

Usage

Run the scripts in the following order:

1. person_detect.rb (optional) to detect and crop people
2. detect.py with bibs
3. person_aggregate.py (optional) to aggregate results from cropped people
4. detect.py with text
5. recognise.rb to recognise the text

Pipeline

To run the pipeline:

make run \
  JOB_ID=unique_job_id
  IN_DIR=/path/to/input/images \
  OUT_DIR=/path/to/output \
  DARKNET_DIR=/path/to/darknet \
  PICKLE_CONFIG_BIB=/path/to/bib/models/config.pickle \
  PICKLE_CONFIG_TXT=/path/to/text/models/config.pickle \
  TESSERACT_BIN_DIR=/path/to/tesseract \
  TESSDATA_DIR=/path/to/tessdata \
  CROP_PEOPLE=1

Use CROP_PEOPLE=0 to not crop people using YOLO.

Output

Sample annotated image:

Each annotation on the image is in the format string_read [bib_accuracy% / txt_accuracy %].

Respective JSON file for this image:

{
   "char":{
      "regions":[
         {
            "y2":641,
            "width":62,
            "height":37,
            "char":"~",
            "x2":477,
            "y1":567,
            "x1":353
         }
      ],
      "string":"~",
      "elapsed_seconds":0.338908
   },
   "person":{
      "regions":[
         {
            "y1":305,
            "x2":488,
            "x1":152,
            "y2":813,
            "accuracy":0.68619
         },
         {
            "y1":256,
            "x2":829,
            "x1":285,
            "y2":863,
            "accuracy":0.676243
         },
         {
            "y1":246,
            "x2":1016,
            "x1":615,
            "y2":841,
            "accuracy":0.659793
         }
      ],
      "elapsed_seconds":1.251645
   },
   "bib":{
      "regions":[
         {
            "y1":403,
            "x2":874,
            "x1":705,
            "y2":540,
            "accuracy":0.9993394017219543
         },
         {
            "y1":523,
            "x2":452,
            "x1":312,
            "y2":641,
            "accuracy":0.9995879530906677
         },
         {
            "y2":453,
            "height":36,
            "x2":470,
            "width":36,
            "y1":407,
            "x1":424,
            "accuracy":0.8520742654800415
         },
         {
            "y2":384,
            "height":123,
            "x2":699,
            "width":150,
            "y1":256,
            "x1":511,
            "accuracy":0.8540436625480652
         },
         {
            "y2":733,
            "height":164,
            "x2":666,
            "width":218,
            "y1":515,
            "x1":394,
            "accuracy":0.9998284578323364
         }
      ],
      "elapsed_seconds":29.776029348373413
   },
   "text":{
      "regions":[
         {
            "y2":604,
            "y1":567,
            "height":25,
            "width":50,
            "x2":415,
            "x1":353,
            "accuracy":0.9343273043632507
         }
      ],
      "elapsed_seconds":9.258251905441284
   }
}

The elapsed_time field is in seconds.