Accepted to ICCV 2025 Conference

Lift to Match (L2M) is a two-stage framework for dense feature matching that lifts 2D images into 3D space to enhance feature generalization and robustness. Unlike traditional methods that depend on multi-view image pairs, L2M is trained on large-scale, diverse single-view image collections.

• Stage 1: Learn a 3D-aware ViT-based encoder using multi-view image synthesis and 3D Gaussian feature representation.
• Stage 2: Learn a feature decoder through novel-view rendering and synthetic data, enabling robust matching across diverse scenarios.

🚧 Code under construction.

We compare the 3D-aware ViT encoder from L2M (Stage 1) with other recent methods:

• DINOv2
• FIT3D
• Ours: L2M Encoder

Below are feature comparison results on the Sacré-Cœur dataset:

To get the results, make sure your checkpoints and image files are in the correct paths, then run:

python vis_feats.py \
  --img_paths assets/sacre_coeur_A.jpg assets/sacre_coeur_B.jpg \
  --ckpt_dino ckpts/dinov2.pth \
  --ckpt_fit3d ckpts/fit3d.pth \
  --ckpt_L2M ckpts/output_20250629/l2m_vit_base.pth \
  --save_dir outputs_vis_feat

We synthesize novel-view images with dense matching labels from single-view inputs for training.
Scripts for data generation will be released soon.

We build upon recent advances in ROMA and FIT3D.