Datasets

This is where you keep training data. A dataset is usable as soon as you create it, and you can add or delete files afterwards. When you upload files, the server validates each one: it determines the kind (image, annotation, point cloud, time series, tabular), pairs images with annotations, and automatically updates the dataset summary (manifest).

Every member can view; creating, uploading files and deleting need dataset_write (DEVELOPER, OPERATOR, ADMIN).

Dataset list#

The table columns are name, source, modality, format, records, size and validation. Modality (모달리티) is the kind of data (point cloud, image, time series, tabular, mixed), and Source (소스) is a note on where the data came from. Datasets linked to an external source get a Sync (동기화) badge in the source column. Click the name, records or size header to sort. Click a row to open the detail.

New dataset#

Choose the method with the two cards at the top of the dialog.

Direct upload (직접 업로드) — set only the name, modality and source (optional) to create an empty dataset, then upload files on the detail screen. Time series and tabular data are both CSV/Parquet, so the modality you choose here sets the kind (time series is a measurement history with a time axis; tabular is one row per subject).

Sync from an external source (외부 소스에서 동기화) — choose a remote dataset on a registered external server to create the dataset, and the first sync starts right away. Leave the name empty to use the remote name. The modality defaults to the value the remote reports; if the remote reports none, CSV/Parquet is treated as time series, so pick tabular yourself for tabular data. A remote dataset already linked to another dataset cannot be chosen; press Sync (동기화) on that dataset instead.

The external source list (the External sources (외부 소스) button) shows the registered servers and the number of linked datasets. Every member can see the list; Register source (소스 등록), connection check, edit and delete need dataset_write. Click the linked-dataset count to filter the dataset list to that source's datasets; Clear filter (필터 해제) clears it. Deleting a source keeps the files already fetched. The Integration guide covers registering an external source and syncing.

Dataset detail — Overview#

Further down come the validation results (검증 결과) (schema conformance, label match rate and so on) and the training usage history (학습 사용 이력). A dataset that arrived through an external manifest also gets a variable spec (변수 사양) table (variable, type, unit, mean, missing rate) next to the validation results. If the dataset has PLY files that passed validation, a Point label reading (포인트 라벨 해석) card appears between the file summary and the manifest.

Training usage history#

Each row is one training run that used this dataset, most recently finished first. The server records a row automatically when a training run that reached running (RUNNING) ends (completed, failed, stopped or interrupted). A run that failed during preparation leaves no row. With no history, "No usage history (사용 이력이 없습니다.)" appears.

Deleting the training run record keeps this history. A dataset with history has deletion limits (Deletion limits).

Point label reading (PLY)#

Besides coordinates and color, the points in a PLY file carry integer properties. This card decides which property is the class (what the point is) and which is the instance ID (which object it belongs to). The instance ID is not used as a training class. Training, preview and the wizard all follow this reading.

If no one has set it, the server detects it from the property names. A name containing inst · object is read as the instance ID; a name containing label · class · segment · semantic is read as the class. While the reading is auto-detected, a banner "PLY label properties were detected by name (PLY 라벨 속성을 이름으로 자동 판정했습니다)" above the tabs shows the result. If it is right, press Confirm as is (이대로 확정); if not, press Change (변경) to go to the card and fix it.

Press Save and revalidate (저장하고 다시 검증) to save the reading and revalidate every point cloud file. The original files do not change. If the dataset has training usage history, the confirmation also warns "N runs trained on this dataset used the previous reading (이 데이터셋으로 학습한 Run N개는 이전 해석으로 학습되었습니다)". Reset to auto-detect (자동 판정으로 되돌리기) deletes the saved reading. Editing needs dataset_write; other members only see the reading that is set. It does not save while any file is syncing or being validated.

Enter Background instance IDs (배경 인스턴스 ID) separated by commas (for example 0, up to 32). Points with these IDs are left out of the instance list and assignment. Negative IDs are always background. If you choose a Background class (배경 점의 클래스), the background also trains as that class. Instance segmentation does not detect the background class as objects. With None (ignored in training) (없음 (학습에서 무시)), background points are not used in training.

Renaming a class changes the assignments in every file. Deleting a class used in assignments asks "Unassign and delete? (할당을 해제하고 지울까요?)"; press Unassign and delete (해제하고 지우기) and those instances become unassigned. Switching the class source to From property (속성에서) does not delete assignments; switch back to Assign per instance (인스턴스에 할당) and they are used as before.

The steps for attaching classes to instances are in Assign classes to point cloud instances.

Dataset detail — Files#

In the table view, select several files and delete them at once with Delete selected (선택 삭제) (dataset_write; locked during a sync). The selection lasts only within the page you are viewing. The Summary (요약) column shows a per-file summary such as size, labels and match rate, or the error message for an error file. While a sync runs, a progress banner with Cancel sync (동기화 취소) appears at the top of the Files tab.

File preview#

The preview depends on the file kind. Move through the list with the previous and next arrows at the top, and use Download (다운로드) and Delete (삭제) (dataset_write; deletes at once without a confirmation). Next to the title are the kind and status badges and an origin badge (Direct upload (직접 업로드) or External sync (외부 동기화)). A synced file deleted here is fetched again on the next sync if it still exists remotely.

PLY 3D viewer:

If the file has both classes and instances, a Color by (색 기준) toggle (Classes (클래스) · Instance (인스턴스)) appears above the legend. In the instance view, the legend shows the number of instances and the grey Unassigned (미할당) · Background (배경) point counts. If the reading is auto-detected, the legend carries an Auto-detected (자동 판정) badge.

Time-series CSV:

Statuses and badges#

Dataset validation badge:

File status:

External sync statuses are Pending · Syncing · Completed · Failed · Canceled (대기 · 동기화 중 · 완료 · 실패 · 취소됨).

Deletion limits#

Deleting a dataset and deleting files (one file or Delete selected (선택 삭제)) first check the training runs that use this dataset. The check is per dataset.

A run that reached running leaves history even when it fails or is stopped, so a dataset used by such a run needs ADMIN even to delete a single file. Deleting the dataset also removes its usage history.