Summary: What You’ve Learned#

You’ve completed the Annotation Getting Started Guide. Here’s what you can now do.

Quickstart Track#

You learned the basics of multimodal annotation:

  • Load grouped datasets with synchronized 2D/3D data

  • Switch between camera slices and point clouds

  • Enter Annotate mode (sample modal -> Annotate tab)

  • Create label fields with enforced schemas

  • Draw bounding boxes on images

  • Explore point cloud data in the 3D visualizer

Full Loop Track#

You built a complete data-centric annotation pipeline for multimodal data:

Step 2: Setup Splits

Cloned quickstart-groups to annotation_tutorial. Created test (frozen), val (iteration), golden (QA), and pool splits at the group level to prevent data leakage across modalities.

Step 3: Smart Selection

Used ZCore diversity scoring on camera images to select high-coverage scenes. Better than random.

Step 4: 2D Annotation

Labeled detections on left camera images with KITTI schema enforcement. Only samples with actual labels get marked as annotated.

Step 5: 3D Annotation

Annotated cuboids on point clouds using transform controls. Used camera projections to verify 3D→2D alignment.

Step 6: Train + Evaluate

Trained YOLOv8 on camera images, evaluated on val set, tagged FP/FN failures for targeting.

Step 7: Iteration

Ran Golden QA check, then selected next batch using hybrid strategy: 30% coverage + 70% targeted.

Key Takeaways#

  1. Group-level splits are non-negotiable. Without them, the same scene leaks between train and test.

  2. Label smarter, not harder. Diversity sampling + failure targeting beats random selection.

  3. 30% coverage budget matters. Only chasing failures creates a model that fails on normal cases.

  4. Cross-modal consistency. 2D and 3D labels should agree on the same objects.

  5. QA before training. Golden QA checks catch annotation drift early.

  6. Understand your failures. FP/FN analysis tells you what to label next.

When to Use What#

In-app annotation is good for:

  • Small to medium tasks (tens to hundreds of samples)

  • Multimodal data with linked 2D/3D views

  • Quick corrections and QA passes

  • Prototyping label schemas

  • Single annotator workflows

  • Tight model-labeling feedback loops

Use external tools (CVAT, Label Studio, Labelbox) when:

  • High-volume annotation with distributed teams

  • Complex labeling pipelines requiring external infrastructure

FiftyOne integrates with external annotation tools. See Integrations for details.

What’s Next#

  • Apply to your data - Use this workflow on your production datasets

  • Explore 3D fully - Try more complex 3D scenes with occlusions

  • Scale with teams - The schema and QA workflow supports multiple annotators

  • Explore plugins - Check @voxel51/brain for advanced selection operators

Resources#

Feedback#

Questions or suggestions? Reach us at support@voxel51.com or join our Discord.