MMDetection
MMDetection is an OpenMMLab toolbox for configuring, training and evaluating object detection and related instance recognition models. The competence is navigating its configuration and data pipeline, selecting compatible dependencies and checking annotations and metrics. It enables reproducible experiments across model families rather than representing one particular detection algorithm.
What it is
The toolbox provides modular model components, dataset handling, preprocessing, training and evaluation through the OpenMMLab ecosystem. Configuration specifies the backbone, prediction heads, transforms, optimizer and runtime. A pretrained checkpoint must match the model definition and class inventory, while custom data must satisfy the selected dataset interface. Boxes and masks use explicit coordinate and encoding conventions. Training resumption restores run state, whereas initialization from weights serves a different purpose. Competence includes reading inherited configuration and understanding the effective settings after overrides. Adjacent OpenMMLab packages handle other perception tasks, so functionality should be attributed to the actual toolbox and selected model rather than the ecosystem's collective scope.
What the work involves
Check MMDetection, MMEngine, MMCV and accelerator compatibility for the chosen version. Load a known configuration, then visualize custom annotations as the pipeline reads them. Update class metadata and heads consistently, audit resize and augmentation logic and choose the evaluator for the intended output. Hold out complete scenes or sequences and inspect class-specific recognition errors. Verify effective batch, learning-rate settings and the difference between resume and initialization. Save the resolved configuration and test the checkpoint in a fresh inference process. The deliverable is a reproducible experiment whose dataset, model and metric conventions can be inspected.
Illustrative example
An illustrative project trains a detector for industrial parts using MMDetection. The engineer converts annotations into the expected format and checks overlays after augmentation. A copied configuration still contains the original class count, so it is corrected before training. Held-out images come from separate capture sessions. A small interrupted run validates checkpoint resumption, and the final evaluation checks small fasteners separately from larger parts. Inference inspection confirms correct labels and coordinates on original-resolution photographs.
Limits and common mistakes
Inherited configuration can hide assumptions about batch size, labels or preprocessing. Binary extension and dependency mismatches can prevent execution. A model-zoo result does not establish performance on custom categories. Incomplete annotations and mismatched category IDs can corrupt evaluation without obvious runtime errors. The toolbox's modularity does not remove the need to understand the selected detector. Pin dependencies, inspect the resolved pipeline and validate complete saving and inference behavior, including deployment format constraints.
Prerequisites
Related skills
- → is an instance of: Object Detection
Sources and further reading
- MMDetection Documentation
Toolbox architecture, supported recognition workflows and ecosystem dependencies.
- MMDetection: Train predefined models on standard datasets
Configuration, custom dataset training, checkpoint resumption and evaluation procedures.
Last updated: 2026-10-10