2. Create a dataset
Create a dataset and upload image and annotation files
Required permission: Write datasetsDEVELOPER · OPERATOR · ADMIN
A dataset is created right away from just a name, and afterwards you can add or remove as many files as you like on its detail screen. There is no separate "confirm" step. Each time files arrive, the server validates them and rewrites the file list (manifest) and the counts.
In this tutorial you upload 8 pairs of field camera photos (.jpg) and LabelMe-format annotations (.json). There are two classes: helmet (safety helmet) and head (bare head).
- In the sidebar, open Datasets (데이터셋) and press New dataset (새 데이터셋) ①.
Dataset list — ① New dataset - Choose Direct upload (직접 업로드) and enter a Name (이름). Set Modality (모달리티) to Images (+annotations) (이미지 (+어노테이션)), and use Source (optional) (소스 (선택)) to note where the data came from. Press Register (등록) to go to the dataset detail.
New dataset — direct upload, name, modality and source - On the Files (파일) tab, drag files or folders onto the dashed box ① (you can also click it to choose). Each image is paired with the annotation of the same name, and Valid (유효) or Error (오류) appears as each file uploads.
Files tab of the dataset — ① drop zone ② Train with this dataset - On the Overview (개요) tab, check File summary (파일 집계) and Validation results (검증 결과). When you see "Every image has a matching label (모든 이미지에 라벨 짝이 있습니다)" and the Trainable (학습 가능) badge next to the name, the dataset can be used for training.
Dataset overview — validation results and class distribution update as soon as files arrive
Files you can upload
| Modality | Extensions |
|---|---|
| Images (+annotations) | .png .jpg .jpeg .bmp .tiff .webp + .json .xml, bundles as .zip |
| Point cloud (PLY) | .ply, .zip |
| Time series / tabular | .csv .tsv .parquet, .zip |
- The server recognizes LabelMe, COCO and VOC annotations during validation.
- Files over 32 MiB and
.zipfiles are split into pieces (chunks) by the browser automatically, so an interrupted upload resumes. A single file can be up to 2 GiB (server default). - Only the file name is kept, without the folder structure. Be careful: files with the same name in different folders overwrite each other.
Press Train with this dataset (이 데이터셋으로 학습) ② at the top right (in the Files tab screenshot) to open the new training wizard with this dataset already selected.
Next: 3. New training