Managing Tasks

Tasks are the core annotation units in CVAT. Each task contains uploaded data (images or videos) that will be annotated according to the project’s label schema.

Task Lifecycle

A task progresses through several stages:

Creation: Task is created with configuration
Data upload: Images/videos are uploaded to the task
Job generation: Task is split into jobs for annotators
Annotation: Annotators label the data
Validation: Quality review and corrections
Acceptance: Final approval
Completion: Task marked as complete

Creating a Task

Using the Web UI

Navigate to Projects → Select your project
Click Create a new task
Configure task settings:
- Name: Descriptive task name
- Subset: Optional subset identifier (e.g., “train”, “val”, “test”)
- Advanced configuration:
  - Overlap: Number of frames overlapping between jobs
  - Segment size: Frames per job (0 = single job)
  - Image quality: Compressed frame quality (0-100)
  - Data chunk size: Frames per chunk for streaming
Upload data (see Uploading Data)
Click Submit

Using the REST API

Create a task with POST /api/tasks:

curl -X POST "https://app.cvat.ai/api/tasks" \
  -H "Authorization: Token <your-token>" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Traffic Scene Batch 1",
    "project_id": 1,
    "overlap": 5,
    "segment_size": 100,
    "subset": "train"
  }'

Response (201 Created):

{
  "id": 10,
  "name": "Traffic Scene Batch 1",
  "project_id": 1,
  "owner": {
    "id": 1,
    "username": "admin"
  },
  "assignee": null,
  "status": "annotation",
  "overlap": 5,
  "segment_size": 100,
  "subset": "train",
  "mode": "annotation",
  "dimension": "2d",
  "size": 0,
  "created_date": "2026-03-04T10:30:00.000Z",
  "updated_date": "2026-03-04T10:30:00.000Z",
  "jobs": {
    "count": 0,
    "completed": 0,
    "validation": 0
  }
}

Using the Python SDK

from cvat_sdk import Client, models

client = Client(url="https://app.cvat.ai")
client.login(("username", "password"))

# Create task
task = client.tasks.create(
    spec=models.TaskWriteRequest(
        name="Street Scene Annotations",
        project_id=1,
        overlap=5,
        segment_size=100,
        subset="validation"
    )
)

print(f"Created task ID: {task.id}")

Using the CLI

# Create task in a project
cvat-cli --auth user:password task create \
  --project_id 1 \
  --overlap 5 \
  --segment_size 100 \
  --subset train \
  "Pedestrian Detection Task 1"

# Create standalone task with labels
cvat-cli --auth user:password task create \
  --labels '[{"name": "person"}, {"name": "bike"}]' \
  --overlap 0 \
  "Independent Task"

Uploading Data

Local Files

Upload images or videos from your computer: Using the SDK:

from cvat_sdk.core.proxies.tasks import ResourceType

# Upload local images
task.upload_data(
    resources=[
        "path/to/image1.jpg",
        "path/to/image2.jpg",
        "path/to/image3.jpg"
    ],
    resource_type=ResourceType.LOCAL,
    params={
        "image_quality": 70,
        "chunk_size": 72,
        "sorting_method": "natural"
    }
)

# Upload video
task.upload_data(
    resources=["path/to/video.mp4"],
    resource_type=ResourceType.LOCAL,
    params={
        "frame_filter": "step=5",  # Every 5th frame
        "start_frame": 0,
        "stop_frame": 1000
    }
)

Using the API:

curl -X POST "https://app.cvat.ai/api/tasks/{task_id}/data" \
  -H "Authorization: Token <your-token>" \
  -F "image_quality=70" \
  -F "client_files[0][email protected]" \
  -F "client_files[1][email protected]" \
  -F "client_files[2][email protected]"

Server Files

Use files from a mounted share:

task.upload_data(
    resources=[
        "/mnt/share/dataset/image1.jpg",
        "/mnt/share/dataset/image2.jpg"
    ],
    resource_type=ResourceType.SHARE,
    params={"copy_data": True}
)

Remote Files (URLs)

Provide URLs to download:

task.upload_data(
    resources=[
        "https://example.com/image1.jpg",
        "https://example.com/image2.jpg"
    ],
    resource_type=ResourceType.REMOTE
)

Cloud Storage

Use files from cloud storage:

task.upload_data(
    resources=["manifest.jsonl"],
    resource_type=ResourceType.SHARE,
    params={
        "cloud_storage_id": 5,
        "filename_pattern": "*.jpg"
    }
)

Data Upload Parameters

Parameter	Type	Description
`image_quality`	integer	Compression quality 0-100 (default: 70)
`chunk_size`	integer	Frames per chunk (default: 72)
`start_frame`	integer	First frame to include (default: 0)
`stop_frame`	integer	Last frame to include
`frame_filter`	string	Frame filter: `step=N` to skip frames
`sorting_method`	string	`lexicographical`, `natural`, `predefined`, `random`
`use_cache`	boolean	Enable server-side caching
`copy_data`	boolean	Copy data from share to CVAT

Job Management

Understanding Jobs

When data is uploaded, CVAT automatically splits the task into jobs based on segment_size. Jobs are assigned to annotators for parallel work. Job types:

annotation: Standard annotation job
ground_truth: Ground truth validation job
consensus_replica: Consensus annotation job

Viewing Jobs

# Get all jobs for a task
jobs = task.get_jobs()

for job in jobs:
    print(f"Job {job.id}: frames {job.start_frame}-{job.stop_frame}")
    print(f"  Status: {job.status}")
    print(f"  Stage: {job.stage}")
    print(f"  State: {job.state}")
    print(f"  Assignee: {job.assignee}")

Assigning Jobs

Using the API:

curl -X PATCH "https://app.cvat.ai/api/jobs/{job_id}" \
  -H "Authorization: Token <your-token>" \
  -H "Content-Type: application/json" \
  -d '{"assignee_id": 5}'

Using the SDK:

job = client.jobs.retrieve(job_id=15)
job.update(models.PatchedJobWriteRequest(assignee_id=5))

Job States and Stages

Stages (workflow phase):

annotation: Initial annotation phase
validation: Review and quality check phase
acceptance: Final approval phase

States (completion status):

new: Not started
in progress: Work in progress
completed: Finished
rejected: Needs rework

# Move job to validation stage
job.update(models.PatchedJobWriteRequest(
    stage="validation",
    state="new"
))

# Mark job as completed
job.update(models.PatchedJobWriteRequest(
    state="completed"
))

Task States

Tasks inherit their status from constituent jobs:

annotation: Jobs are being annotated
validation: Jobs are in validation/review
completed: All jobs completed

Deprecated status field is computed from stage and state:

# Get task status
task_info = client.tasks.retrieve(task_id=10)
print(f"Status: {task_info.status}")
print(f"Jobs: {task_info.jobs['count']} total, {task_info.jobs['completed']} completed")

Validation Modes

CVAT supports different validation configurations:

Ground Truth Validation

Create a ground truth job for quality assessment:

# Create ground truth job
gt_job = client.jobs.create(
    spec=models.JobWriteRequest(
        task_id=10,
        type="ground_truth",
        frame_selection_method="random_uniform",
        frame_count=50,  # 50 random frames
        random_seed=42
    )
)

Frame selection methods:

random_uniform: Random frames across entire task
random_per_job: Random frames from each job
manual: Manually specified frame list

# Manual frame selection
gt_job = client.jobs.create(
    spec=models.JobWriteRequest(
        task_id=10,
        type="ground_truth",
        frame_selection_method="manual",
        frames=[0, 10, 20, 30, 40]  # Specific frames
    )
)

Honeypot Validation

Create tasks with hidden validation frames:

# Upload data with validation params
task.upload_data(
    resources=images,
    params={
        "validation_params": {
            "mode": "gt_pool",
            "frame_selection_method": "random_per_job",
            "frames_per_job_count": 10
        }
    }
)

Consensus Annotation

Enable multiple annotators for the same data:

# Create task with consensus replicas
task = client.tasks.create(
    spec=models.TaskWriteRequest(
        name="Consensus Task",
        project_id=1,
        consensus_replicas=3  # 3 annotators per job
    )
)

This creates 3x jobs, allowing agreement analysis.

Updating Tasks

Update Task Properties

task.update(models.PatchedTaskWriteRequest(
    name="Updated Task Name",
    assignee_id=10,
    subset="test"
))

Update Task Labels (Standalone Tasks)

For tasks not in a project:

task.update(models.PatchedTaskWriteRequest(
    labels=[
        models.PatchedLabelRequest(
            id=5,
            name="updated_label_name"
        ),
        models.PatchedLabelRequest(
            name="new_label",
            color="#ff00ff"
        )
    ]
))

Tasks in a project inherit labels from the project. Update labels at the project level.

Deleting Tasks

# REST API
curl -X DELETE "https://app.cvat.ai/api/tasks/{task_id}" \
  -H "Authorization: Token <your-token>"

# Python SDK
task = client.tasks.retrieve(task_id=10)
task.remove()

# Or directly
client.tasks.remove_by_id(10)

# CLI
cvat-cli --auth user:password task delete 10

Best Practices

Optimize job segmentation

Set segment_size to 100-500 frames per job
Use overlap (5-10 frames) for video tasks to ensure continuity
Smaller jobs = easier task management and recovery
Consider annotator capacity when sizing jobs

Configure data upload efficiently

Use chunk_size=72 (default) for optimal streaming
Set image_quality based on needs: 70-85 for most tasks
Use frame_filter="step=N" to reduce video frames
Enable use_cache for faster repeated access

Manage task organization

Use subset field for train/val/test splits
Assign descriptive names with batch/date info
Group related tasks in projects
Assign tasks to appropriate team members

Monitor task progress

Check jobs.completed count regularly
Review job states and stages
Use quality reports to track annotation quality
Set up validation workflows early

Next Steps

Quality Control

Set up validation and quality metrics

Annotation Guide

Learn about annotation tools and workflows

Export Annotations

Export annotated data in various formats

API Reference

Complete task API documentation

Get Started

Annotation

Projects & Tasks

Dataset Management

Integrations

Account & Organization

Managing Tasks

Task Lifecycle

Creating a Task

Using the Web UI

Using the REST API

Using the Python SDK

Using the CLI

Uploading Data

Local Files

Server Files

Remote Files (URLs)

Cloud Storage

Data Upload Parameters

Job Management

Understanding Jobs

Viewing Jobs

Assigning Jobs

Job States and Stages

Task States

Validation Modes

Ground Truth Validation

Honeypot Validation

Consensus Annotation

Updating Tasks

Update Task Properties

Update Task Labels (Standalone Tasks)

Deleting Tasks

Best Practices

Next Steps

Quality Control

Annotation Guide

Export Annotations

API Reference

Build docs developers (and LLMs) love

Get Started

Annotation

Projects & Tasks

Dataset Management

Integrations

Account & Organization

​Task Lifecycle

​Creating a Task

​Using the Web UI

​Using the REST API

​Using the Python SDK

​Using the CLI

​Uploading Data

​Local Files

​Server Files

​Remote Files (URLs)

​Cloud Storage

​Data Upload Parameters

​Job Management

​Understanding Jobs

​Viewing Jobs

​Assigning Jobs

​Job States and Stages

​Task States

​Validation Modes

​Ground Truth Validation

​Honeypot Validation

​Consensus Annotation

​Updating Tasks

​Update Task Properties

​Update Task Labels (Standalone Tasks)

​Deleting Tasks

​Best Practices

​Next Steps

Quality Control

Annotation Guide

Export Annotations

API Reference

Build docs developers (and LLMs) love

Task Lifecycle

Creating a Task

Using the Web UI

Using the REST API

Using the Python SDK

Using the CLI

Uploading Data

Local Files

Server Files

Remote Files (URLs)

Cloud Storage

Data Upload Parameters

Job Management

Understanding Jobs

Viewing Jobs

Assigning Jobs

Job States and Stages

Task States

Validation Modes

Ground Truth Validation

Honeypot Validation

Consensus Annotation

Updating Tasks

Update Task Properties

Update Task Labels (Standalone Tasks)

Deleting Tasks

Best Practices

Next Steps