Model description

RoMa is a state of the art deep keypoint detector and matcher. It falls in the "fully dense" matcher category, contrary to semi-dense matchers like LoFTR or Sparse matchers like Superpoint+SuperGlue.

When you want the best performance when you do not have many limitation from speed or memory, then RoMa is giving among the best results.

The paper introduces several key contributions:

• Feature Integration: Combines DINOv2’s coarse features with fine ConvNet features for robust, precise localization.
• Transformer-Based Match Decoder: Predicts anchor probabilities instead of coordinates.
• Improved Loss Formulation: Uses classification-based regression for global matches and robust regression for refinement.

Open source status

• The model implementation is available
• The model weights are available

Provide useful links for the implementation

Paper : https://arxiv.org/abs/2305.15404
Code : https://github.com/Parskatt/RoMa