🔍After Annotation

You've finished annotating your dataset! 🎉

Annotation Finished
YAY!

A high-quality dataset is consistent. You may easily follow this document to perform a quick audit of your annotations before moving to training.

Quick-Check via Dataset Filters

In the Image Annotation Window, use the filter dropdown to isolate specific labeling states.

Filter
Logic
What to look for

All

Total dataset

General overview of project volume.

Annotated

≥1 Bounding Box

Ensure boxes are tight and classes are correct.

Empty

Background/Negative

Critical: Confirm these truly have no objects.

Excluded

No annotation file

Ensure no usable data was accidentally hidden.

Common Pitfall: Having "Empty" images that actually contain objects will confuse the model. If an object is there, it must be labeled or the image must be "Excluded."

Quick Review: Keyboard Shortcuts

These shortcuts allow for rapid auditing without leaving the canvas.

Navigation & Class Selection

  • D / A: Next / Previous image.

  • Shift + D / Shift + A: Jump forward/back by 10 images.

  • S / W: Next / Previous class.

  • Shift + S / Shift + W: Jump classes by 3.

  • H (Hold): Temporarily hide annotations to see the raw image.

Labeling & File Management

  • O: Mark as Background (Creates/clears an empty annotation file).

  • P: Exclude Image (Removes the annotation file).

  • X: Remove last bounding box.

  • Shift + C: Clear all boxes on the current image.

  • M: Move image + annotation to a /moved subfolder (Folder Mode).

  • Shift + Delete: Permanently delete image + annotation.

Advanced Analysis Tools

A) Class Frequency Analysis

Open Tools → Class Frequency Analysis to visualize your data distribution.

  • Rare Classes: If a class is significantly lower than others, the model may ignore it.

  • Dominant Classes: If one class makes up the bulk of the data, the model may over-predict it.

If you find imbalances, consider collecting more data for rare classes or removing some examples of dominant classes with redundant images.

Another option is to use data augmentation techniques to artificially increase the variety of underrepresented classes.

B) Pattern Recognition

Watch for these "Annotation Quality" issues during your review:

  • Loose Boxes: Too much background noise inside the box.

  • Inconsistent Style: Mixing tight and loose boxes across the same class.

  • Missing Negatives: Not enough "Empty" images to teach the model what isn't an object.

Audit Routine

  1. Filter to Annotated: Review ~20–50 images across the entire set (not just the first page).

  2. Filter to Empty: Review ~10–20 images to ensure they are truly empty.

  3. Spot-check Excluded: Ensure no high-quality data is sitting idle.

  4. Edge-Case Pass: Search for the smallest objects, worst glare, and heaviest motion blur.

Validation Set

Pick 30–100 images or video clips that represent "Real World" challenges (bad lighting, clutter, etc.).

Keep these labeled perfectly. Use this set as your final "reality check" before deploying any model to production.

Backup Reminder: Always duplicate your dataset folder before running batch operations or mass-deletions.

PreviousAugment DatasetNextTrain Object Detection(YOLO) Models

Last updated

Was this helpful?