PlaneCycle

PlaneCycle: Training-Free 2D-to-3D Lifting of Foundation Models Without Adapters (arXiv)

Overview

PCA visualizations of frozen lifted DINOv3 features
On three 3D datasets across HW, DW, and DH planes (inconsistencies circled).

Overview of PlaneCycle across three orthogonal planes (HW, DW, DH).
Flattened slice tokens are processed by shared ViT layers with plane-specific RoPE.

Code structure

• planecycle Contains the foundational logic for the PlaneCycle architecture.
  • operators: Implementation of PlaneCycle operator.
  • converters: ViT converters.
• experiments/dinov3 the scripts to run experiments.
  • medminist: Comprehensive pipelines for training and benchmarking on the 6 different 3D MedMNIST+ datasets.
• dinov3/dinov3/models
  • vision_transformer: Our modified Vision Transformer implementation.