YOLOv5 with LMDB support

This fork was created in order to add support for a specific dataset format to YOLOv5.

LMDB dataset format

Said dataset format consists of a collection of LMDB files containing images and labels; each key contains one image and its labels. Keys can be any string, as long as they are a valid file name. The program may break if keys contain characters not allowed in file paths. Labels are JSON strings with the following structure:

{
  'source': 'Various Placards',
  'file_name': 'file-name.jpg',
  'shape': [600, 400, 3],
  'boxes': [
    [ 0.6638655462184874,
      0.4476987447698745,
      0.06554621848739496,
      0.0899581589958159,
      7],
    [ 0.7294117647058823,
      0.45188284518828453,
      0.05546218487394958,
      0.08158995815899582,
      12]
  ]
}

The boxes array contains the actual labels and is the only required field. Labels are slightly different from those in the YOLO format. Specifically, they represent:

1. the x coordinate of the top left corner of the box, normalized by the width of the image,
2. the y coordinate of the top left corner of the box, normalized by the height of the image,
3. the width of the box, normalized by the width of the image,
4. the height of the box, normalized by the height of the image,
5. the class (optional, an integer).

Shape values represents image height, width and depth, respectively.

NOTE: image segmentation is not supported.

Converting a YOLO dataset to LMDB

Datasets in the YOLOv5 format can be converted to this LMDB format using the yolo2lmdb.py script.

Enabling LMDB dataset support in YOLOv5

The LMDB support is enabled by setting the lmdb field to true in the dataset's .yaml. The train and val fields must point to the folder containing the keys.json of each datasets. Said fields can also be lists of paths.

lmdb: true
train: 
 - /usr/src/app/AI-Data/VTE-lmdb/Vte_Lmdb_train
 - /usr/src/app/AI-Data/VTE-lmdb/Vte_Lmdb_bg
val: 
 - /usr/src/app/AI-Data/VTE-lmdb/Vte_Lmdb_test
nc: 1
names: ['person']