Overview
Our users’ primary interest is computer vision test and evaluation. This includes exploration of labeled image and video datasets. Fifty One is the most widely recognized open source viewer for vision datasets. It helps users:
Visualize Data: Quickly search, filter, and view massive datasets of images, videos, or 3D point clouds.
Evaluate Models: Overlay your model's predictions (like bounding boxes or segmentation masks) onto ground-truth data to visually and statistically spot where your model is making mistakes.
Curate Datasets: Find duplicate images, compute image embeddings, and slice your dataset to find edge cases (e.g., "show me all images where the model missed a pedestrian at night").
What it does not do is allow users to label datasets or modify existing labels. That functionality requires CVAT or LabelStudio which would be (most likely) in a separate pack and is outside the scope of this pack MVP.
Important: The datasets will be held on the data catalog in a separate pack (see above). It would be nice if we could also explore datasets from the Jupyter storage space, but that might not be a reasonable request.
Requirements
| |
User Requirement |
| 1 |
Must be deployed on AWS GovCloud (COSMOS) |
| 2 |
Must support multi-user (user identities tied back to keycloak) and be performant at scale |
| 3 |
Must support permission control - restricted access to datasets based on user or user group (should be able to rely on the RBAC from the data catalog) |
| 4 |
Must be able to access datasets from data catalog |
| 5 |
Nice to have: access datasets from JupyterHub storage |
Overview
Our users’ primary interest is computer vision test and evaluation. This includes exploration of labeled image and video datasets. Fifty One is the most widely recognized open source viewer for vision datasets. It helps users:
Visualize Data: Quickly search, filter, and view massive datasets of images, videos, or 3D point clouds.
Evaluate Models: Overlay your model's predictions (like bounding boxes or segmentation masks) onto ground-truth data to visually and statistically spot where your model is making mistakes.
Curate Datasets: Find duplicate images, compute image embeddings, and slice your dataset to find edge cases (e.g., "show me all images where the model missed a pedestrian at night").
What it does not do is allow users to label datasets or modify existing labels. That functionality requires CVAT or LabelStudio which would be (most likely) in a separate pack and is outside the scope of this pack MVP.
Important: The datasets will be held on the data catalog in a separate pack (see above). It would be nice if we could also explore datasets from the Jupyter storage space, but that might not be a reasonable request.
Requirements