Data configuration and customization

UP supports two types of datasets: the public dataset and the custom dataset.

Public dataset

Currently, the following public datasets are supported.

CocoDataset

dataset: # Required.
  train:
    dataset:
      type: coco
      kwargs:
        meta_file: coco/annotations/instances_train2017.json
        image_reader:
          type: fs_opencv   # ['fs_opencv', 'fs_pillow', 'ceph_opencv', 'osg']
          kwargs:
            image_dir: coco/train2017
            color_mode: RGB
        transformer: [*flip, *train_resize, *to_tensor, *normalize]
  test:
    dataset:
      type: coco
      kwargs:
        meta_file: &gt_file coco/annotations/instances_val2017.json
        image_reader:
          type: fs_opencv   # ['fs_opencv', 'fs_pillow', 'ceph_opencv', 'osg']
          kwargs:
            image_dir: coco/val2017
            color_mode: RGB
        transformer: [*test_resize, *to_tensor, *normalize]
        evaluator:
          type: COCO               # choices = {'COCO'}
          kwargs:
            gt_file: *gt_file
            iou_types: [bbox]
  batch_sampler:
    type: aspect_ratio_group
    kwargs:
      sampler:
        type: dist
        kwargs: {}
      batch_size: 2
      aspect_grouping: [1,]
  dataloader:
    type: base
    kwargs:
      num_workers: 4
      alignment: 32

• You need to set the ‘meta_file’ and ‘image_dir’ in dataset, and set augmentation in transformer.

• UP seperates the dataset and evalutor for adapting different kinds of datasets and evalutors.

• UP supports two kinds of evalutors for different datasets: coco(CocoEvaluator) - CocoDataset, MREvaluator(CustomDat aset) - CustomDataset.

Training on the custom dataset

Dataset structure

The structure of dataset is as followed.

datasets
├── image_dir
|   ├── train
|   ├── test
|   └── val
└── annotations
    ├── train.json
    ├── test.json
    └── val.json

Annotation file format

We suggest that saving the annotation file with json format. A single image is annotated as followed.

{
    "filename": "000005.jpg",
    "image_height": 375,
    "image_width": 500,
    "instances": [                // List of labeled entities, optional for test
      {
        "is_ignored": false,
        "bbox": [262,210,323,338], // x1,y1,x2,y2
        "label": 9                 // Label id start from 1, total C classes corresponding  a range of [1, 2, ..., C]
      },
      {
        "is_ignored": false,
        "bbox": [164,263,252,371],
        "label": 9
      },
      {
        "is_ignored": false,
        "bbox": [4,243,66,373],
        "label": 9
      }
    ]
}

Accuracy evalution

Config is as followed.

evaluator:
  type: MR
  kwargs:
    gt_file: path/your/test.json
    iou_thresh: 0.5
    num_classes: *num_classes
    data_respective: False

Note

• UP supports multiple training/testing evalutions. When ‘data_respective’ is set to True,the datasets will be respectively evaluted; when it is set to False, the datasets will be together evaluted.

• When ‘data_respective’ is set to True, we suggest setting the batchsize to 1, or the padding between different data will effect the final accuracy.

UP supports evaluting the custom dataset by MR mode which contains two indices as followed.

• MR@FPPI=xxx: Miss rate while ‘FPPI’ reaches some value.

• Score@FPPI=xxx: Confidence score while ‘FPPI’ reaches some value.

Illustration of the config

dataset:
  train:
    dataset:
      type: custom
      kwargs:
        num_classes: *num_classes
        meta_file: # fill in your own train annotation file
        image_reader:
          type: fs_opencv
          kwargs:
            image_dir: # fill in your own train data path
            color_mode: RGB
        transformer: [*flip, *resize, *to_tensor, *normalize]
  test:
    dataset:
      type: custom
      kwargs:
        num_classes: *num_classes
        meta_file: # fill in your own test annotation file
        image_reader:
          type: fs_opencv
          kwargs:
            image_dir: # fill in your own test data path
            color_mode: RGB
        transformer: [*resize, *to_tensor, *normalize]
        evaluator:
          type: MR # fill in your own evaluator
          kwargs:
            gt_file: # fill in your own test annotation file
            iou_thresh: 0.5
            num_classes: *num_classes
  batch_sampler:
    type: aspect_ratio_group
    kwargs:
      sampler:
        type: dist
        kwargs: {}
      batch_size: 2
      aspect_grouping: [1,]
  dataloader:
    type: base
    kwargs:
      num_workers: 4
      alignment: 1

Note

• The types of the dataset and evalution need to be set to custom and MR

• The parameter ‘num_classes’ needs to be set.