Skip to content

[SPIKE] CVAT or LabelStudio to label datasets #651

Description

@jallard-ot-ai

Overview

Our primary users are running test and evaluation on computer vision models and datasets. FiftyOne covers viewing of pre-labeled datasets. We need a tool that will enable users to label new datasets or edit existing ones. The two most popular open source tools are CVAT and LabelStudio. These tools help users:

  • Add annotations/labels to image and video datasets

  • Modify existing labels

  • Export various common annotation formats in support of different models

FiftyOne integrates seamlessly with both CVAT and LabelStudio


Important: They are down selecting vendors for the Data Labeling Challenge this week. The tooling that the vendors require may dictate the choice we make here. 

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 transitive permission control - restricted access to datasets based on user or user group (should be able to rely on the RBAC from the data catalog, and the request to data catalog must come from the user’s claim, not a “global/admin” FiftyOne credential)
4 Must be able to access datasets from data catalog
5 Nice to have: access datasets from JupyterHub storage

Decision Guidance on CVAT and LabelStudio

We have not yet made a decision on which to build. Kim is leaning toward CVAT but if several DLC vendors are going to require LabelStudio, that will affect this decision. 


CVAT (Computer Vision Annotation Tool)

CVAT is designed for velocity and precision on visual tasks. If your team is annotating

thousands of dense images or long video clips, CVAT is built to make that process as fast as

possible.


Pros:

  • Unmatched Video Tooling: CVAT is arguably the best open-source tool for video. Its

built-in frame interpolation means you only annotate keyframes, and the software

automatically calculates the bounding box movements for the frames in between.

  • Deep Vision Features: It offers highly refined tools for complex vision tasks, including

skeletons (pose estimation), 3D cuboids, AI-assisted masking (Segment Anything), and

intelligent scissors.

  • Speed for Power Users: The interface is heavily optimized for keyboard shortcuts,

allowing experienced annotators to fly through datasets without moving their mouse.

  • Generous Open-Source Version: The self-hosted version includes almost all of the core

annotation features completely free.

Cons:

  • Vision Only: It does not support text (NLP), audio, or time-series data.

  • Rigid Interface: The UI is what it is; you cannot customize the layout to fit a niche

workflow or display conditional logic.

  • Steep Initial Learning Curve: For a beginner just trying to draw a few boxes, the sheer

number of tools and settings can be overwhelming.


Label Studio

Label Studio takes a framework approach. Instead of giving you a fixed annotation interface,

it gives you a blank canvas and XML-like tags to build the exact UI your project requires.


Pros:

  • True Multi-Modality: Label Studio shines when your AI pipelines involve multiple data

types. You can use it for text classification, audio transcription, named entity recognition

(NER), and image segmentation—all under the same platform.

  • Endless UI Flexibility: You can design custom layouts. For example, you can create a

workflow that shows an image, asks the annotator to transcribe text found in the image,

and provides a dropdown to classify the sentiment of that text.

  • Strong Enterprise & MLOps Features: It offers robust integration into modern ML

pipelines, handling cloud-native storage easily, and its Enterprise version has highly

polished role-based access control (RBAC), SSO, and reviewer workflows.

Cons:

  • Weak Video Support: Compared to CVAT, Label Studio’s video capabilities are

rudimentary. It treats video mostly as a sequence of extracted frames and lacks the

smooth, native interpolation required for heavy video tracking tasks.

  • Setup Overhead: You have to build (or configure) your labeling interface using their

declarative XML configuration before you can start annotating.

  • Less Refined for Dense Vision: While it handles bounding boxes and polygons fine, its

vision tools can feel slightly clunky compared to CVAT when annotating highly complex,

dense visual scenes.


The Final Verdict

Choose CVAT if: Your core focus is evaluating and testing dedicated computer vision systems. If you need to rigorously review dense visual data, validate bounding box accuracy, or scrutinize video tracking performance frame-by-frame, CVAT’s specialized tools will save your testing teams hundreds of hours.

Choose Label Studio if: Your evaluation workflows span multiple modalities beyond just pure vision. If you need a single platform to review OCR text extraction, audit audio transcriptions, and edit computer vision labels, Label Studio’s flexible interfaces are unbeatable


Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

Fields

Priority

Critical

Start date

Sep 1, 2026

Target date

Sep 4, 2026

Size

S

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions