Sim2Real Object Detection Present

Fine-tuning YOLO with synthetic data from Yrikka to close the gap between simulated and real-world object detection.

This project, in collaboration with Yrikka through the Cornell Tech AI Studio, tackles the synthetic-to-real (sim2real) performance gap in object detection. Using Yrikka's data engine, the team generates and annotates synthetic images for everyday objects—potted plant, chair, cup, vase, and book—then fine-tunes a YOLO model to boost its accuracy on challenging real-world test sets.

The workflow includes annotation correction with CVAT, targeted synthetic data generation for edge cases, and performance evaluation using mAP@50, aiming for at least a 0.10 improvement over the baseline model.

• Platform

  Google Colab / Kaggle / CVAT

• Model

  Ultralytics YOLO (on PyTorch)

• Tools

  scikit‑learn, CVAT, OpenCV, Matplotlib

• Synthetic Data

  Yrikka Data Engine / APEX API

• Goal

  Improve YOLO object detection mAP@50 by ≥ 0.10 using synthetic data

• Referencehttps://www.yrikka.com/
• Github

  Pending Permission to make it public