Demo video of synthetic data generation

Client

Panda GmbH. A joint project on computer vision quality control for industrial connectors.

Challenge

Build a multi-stage computer vision pipeline for quality control of industrial connectors: first locating each connector, then detecting defects and judging how severe they are. Both stages needed large amounts of accurately labelled training images.

Our approach

• Stage 1, object detection: generated synthetic training data from a 3D model based on the connector's CAD data, randomizing rotation, scale, position, lighting and materials, to train an object detection CNN.
• Stage 2, anomaly detection: procedurally generated a range of defects, each automatically annotated with its severity level, to fine-tune the anomaly detection model.
• Built the entire synthetic data generation (SDG) pipeline in Blender, automated through its Python API (bpy).

Deliverables

• An automated synthetic data generation pipeline.
• Annotated synthetic datasets for object detection and for defect-severity anomaly detection.
• This demo video of the generation process.

Tools