• Trained DINOv2_ViTB14 with most of the existing LEGO boxes.
  • Attained an accuracy of 40% on composed segmentation and model ID recognition, yet to improve.