Cvat Push

Push to CVAT for Review

After training analysis identifies uncertain images, push them to CVAT for human annotation or correction.

Pre-Flight Checklist

• [ ] yolo-analyze has been run (check for uncertain_images.txt or analysis output)
• [ ] CVAT configured in yolo-project.yaml
• [ ] CVAT_ACCESS_TOKEN env var is set

Workflow

1. Gather Uncertain Images

Read the analysis output:

cat reports/uncertain_images.txt

Or check experiments/analysis.md for the list of uncertain predictions.

2. Categorize by Priority

Organize images into review batches:

• False negatives (highest priority): Images where the model missed detections but labels exist
• Low confidence (medium priority): Detections with confidence between lowconf and highconf thresholds
• No detection, no label (low priority): Images where neither model nor labels found anything

3. Push to CVAT

# From analysis file (auto-batches into tasks of 50)
yolo-cvat push --from-analysis reports/uncertain_images.txt

# Or manually specify images
yolo-cvat push --images  --labels  --task-name "Review: False Negatives"

4. Generate Report

Write experiments/cvat_push_report.md:

## CVAT Upload Report
- Date: YYYY-MM-DD
- Tasks created: N
- Total images: N
- False negatives: N (Task ID: X)
- Uncertain: N (Task ID: Y)
- CVAT URL: http://localhost:8080/tasks/

Decision Tree

More than 100 images?
├── Yes → Split into batches of 50, create multiple tasks
└── No → Single task

Has existing YOLO labels for these images?
├── Yes → Upload as pre-annotations (saves annotator time)
└── No → Push images only for fresh annotation

False negatives found?
├── Yes → Create separate high-priority task labeled "URGENT: False Negatives"
└── No → Single "Review: Uncertain" task

Guardrails

• NEVER push more than 200 images without asking the user first
• ALWAYS include pre-annotations when available (reduces annotation time)
• Keep a local copy of all images — CVAT is not a backup
• Write the push report before telling the user it's done

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: MacroMan5
• Source: MacroMan5/autotrain-yolo
• License: MIT

Install and usage instructions live in the source repository linked above.