Skip to main content

Building an AI Development Environment

Project Overview

  • Build an environment that automatically labels the objects and features a perception model needs to recognize.

Key Questions

  • Fine-tune a large model to meet the perception requirements.
  • Address class imbalance in the training data.
  • Supplement the training data for rarely seen objects.

Results

Generating training data for perception models without manual annotation

We built a foundation model for monocular camera perception. It supports object, lane, traffic light, traffic sign, and scene recognition. We fine-tuned the model to the perception requirements and built a pipeline that automatically generates training data for the models used at inference time. For cases that occur infrequently, we use data generation to augment the training set and ensure it covers those cases.

Key Points

  • Built a foundation model that covers the required perception tasks.
  • Developed a way to generate additional training data for rarely seen objects.