COCO Style Custom Dataset 만들기

1. fifty one 설치

• fiftyone : 다양한 오픈 데이터셋을 관리할 수 있는 모듈

  $ pip install fiftyone

• fiftyone luunch_app

import fiftyone as fo

# The directory containing the source images
data_path = "./Zzuri/data/"
# The path to the COCO labels JSON file
labels_path = "./Zzuri/labels.json"

# Import the dataset
dataset = fo.Dataset.from_dir(
    dataset_type=fo.types.COCODetectionDataset,
    data_path=data_path,
    labels_path=labels_path,
)
if __name__ == "__main__":
    # Ensures that the App processes are safely launched on Windows
    session = fo.launch_app(dataset)
    session.wait()

• fifty one app을 launch하면 localhost:5151 웹으로 연결되며 dataset을 감상할 수 있다.
• 세상에서 제일 귀여운 고양이🐈 dataset을 감상해보자!
  

2. COCO Style Custom Dataset 만들기

• caffe 기반에서 사용했던 gt label을 coco style로 변환하였다.

import fiftyone as fo
import os

gt_path = "iray/labels/"
image_path = "iray/images/"

gt_files = os.listdir(gt_path)

dataset = fo.Dataset()
# Set default classes
dataset.default_classes = ["Person", "Vehicle", "Animal", "Bike"]
dataset.save()  # must save after edits

# Create samples for your data
samples = []
for i in range(len(gt_files)):
#for i in range(1):
    image_name = image_path+gt_files[i]
    image_name = image_name[:-3] + 'png'
    print(str(i + 1) + ' : ' + image_name)

    gt_file = open(gt_path+gt_files[i])
    sample = fo.Sample(filepath=image_name)
    detections = []

    for line_text  in gt_file:
        tokens = line_text.split()

        label = dataset.default_classes[int(tokens[0])]

        box_width = float(tokens[3])
        box_height = float(tokens[4])
        left_top_x = float(tokens[1]) - box_width/2
        left_top_y = float(tokens[2]) - box_height/2
        bounding_box = [left_top_x, left_top_y, box_width, box_height]

        detections.append(
            fo.Detection(label=label, bounding_box=bounding_box)
        )
        # Store detections in a field name of your choice
        sample["ground_truth"] = fo.Detections(detections=detections)

    samples.append(sample)

# Create dataset
print('add_samples...')
dataset.add_samples(samples)
dataset.save()  # must save after edits

print('export...')
# Export the dataset
dataset.export(
    export_dir="iray/",
    dataset_type=fo.types.COCODetectionDataset,
    label_field="ground_truth",
)

• 아래와 같이 COCO type의 데이터셋이 생성된다.
  

• jq 다운로드
  jq를 이용해서 만들어진 coco labels.json 파일을 보기 편하게 변환할 수 있다.

  $ jq-win64.exe . .\labels.json > labels_cvt.json

{
  "info": {
    "year": "2022",
    "version": "1.0",
    "description": "Exported from FiftyOne",
    "contributor": "hjkim",
    "url": "https://voxel51.com/fiftyone",
    "date_created": "2022-04-11T18:44:23"
  },
  "licenses": [],
  "categories": [
    {
      "id": 0,
      "name": "Person",
      "supercategory": null
    },
    {
      "id": 1,
      "name": "Vehicle",
      "supercategory": null
    },
    {
      "id": 2,
      "name": "Animal",
      "supercategory": null
    },
    {
      "id": 3,
      "name": "Bike",
      "supercategory": null
    }
  ],
  "images": [
    {
      "id": 1,
      "file_name": "MIR1.0_20190814191628SN_TP29_WN1_C1_NKOR_RAYS_PU_FD_OKSH_00448.png",
      "height": 256,
      "width": 512,
      "license": null,
      "coco_url": null
    },
    .
    .
    .