Docs

The manual and guide text from your BadBob app.

Contents

Capture Lab guide

Find where a model struggles, correct the evidence, then choose where it belongs.

Quick start

Start with Full Test, Focused Test, or Free Capture.

Turn the objects, change angles, and try separated objects, tight clusters, and an outlier.

Correct annotations and classes. Set failure status, add a note, and save the capture.

Choose Add to train, Guard Set, Save later, or Hard case.

Capture Lab · Quick start

Annotator guide

Turn captured images into trusted ground truth by checking every object, correcting its box and class, and saving a clean training project.

Quick start

Use the image list and its status badges to find empty, automatic, or already edited examples.

Move and resize annotations, add missing objects, remove false annotations, and assign the correct class.

Resolve missing images, empty annotations, invalid labels, and boxes that extend outside the image.

Merge the reviewed project into a chosen training add-on only when every image is ready.

Annotator · Quick start

Prepare trusted data

Build reviewed datasets and keep evaluation images separate.

Quick start

Import an existing dataset or collect reviewed images. BadBob verifies copies and reuses matching Image Bank images; originals stay unchanged.

Check class mappings, annotations and split assignments. Dry run reports missing data and overlap before import.

Train teaches the model, Val helps evaluate during development, and Test is reserved for final evaluation. Keep these roles separate.

Choose a name, version and optional parent. A published dataset fixes the image revisions; later changes belong in a new version.

Data Studio · Quick start

Prepare and train

Use a published dataset with reviewed annotations. Validation, test and guard images keep their separate roles.

Quick start

Choose the algorithm, iterations and batch size. Automatic iterations are a BadBob estimate. Give the model a name and version.

Choose an output folder outside the project. You can change tabs during training and enable a completion sound.

Evaluate validation or test images. Inspect missed objects, extra detections and wrong classes. Send selected images to Failure Bank for manual review.

Compare models on the same images, with per-class precision and recall. Guards are scored separately. External models are evaluated without resuming their training.

Training · Quick start

Failure Bank guide

Review model failures, preserve regression evidence, and prepare verified training handoffs without changing the source evidence.

Quick start

Inspect the model prediction beside the current human revision while BadBob verifies the selected source image.

Record what happened, why the example matters, and the context another reviewer will need later.

Send useful examples toward training, protect regressions in the Guard Set, or keep uncertain cases for later.

Hold uncertain evidence outside train, validation, and test. Hard cases stay visible for investigation and later curriculum stages.

Failure Bank · Quick start