Implementation:Kornia Kornia LoFTR

Knowledge Sources	LoFTR: Detector-Free Local Feature Matching with Transformers Kornia
Domains	Vision, Feature_Matching, Deep_Learning
Last Updated	2026-02-09 15:00 GMT

Overview

Concrete tool for detector-free dense feature matching using Transformers provided by Kornia.

Description

The LoFTR class implements the Detector-Free Local Feature Matching with Transformers architecture. It accepts a dictionary with two grayscale images {"image0": (B,1,H,W), "image1": (B,1,H,W)} and returns matched keypoint coordinates, confidence scores, and batch indices. Pretrained weights are available for "outdoor" and "indoor" scenes. The module handles feature extraction, coarse matching via Transformer attention, and fine-level sub-pixel refinement.

Usage

Initialize with pretrained weights and pass image pair dictionaries. Commonly used with ImageStitcher or as input to RANSAC for geometric verification.

Code Reference

Source Location

Repository: kornia
File: kornia/feature/loftr/loftr.py
Lines: L70-212

Signature

class LoFTR(nn.Module):
    def __init__(
        self,
        pretrained: Optional[str] = "outdoor",
        config: dict[str, Any] = default_cfg
    ) -> None

Forward

def forward(
    self,
    data: dict[str, torch.Tensor]
) -> dict[str, torch.Tensor]

Import

from kornia.feature import LoFTR

I/O Contract

Inputs

Name	Type	Required	Description
data	dict[str, torch.Tensor]	Yes	Dictionary with keys "image0" and "image1", each a grayscale tensor of shape (B, 1, H, W)

Outputs

Name	Type	Description
keypoints0	torch.Tensor (N, 2)	Matched keypoint coordinates in image0
keypoints1	torch.Tensor (N, 2)	Matched keypoint coordinates in image1
confidence	torch.Tensor (N,)	Confidence scores for each match
batch_indexes	torch.Tensor (N,)	Batch index for each match

Usage Examples

Basic Matching Between Two Images

import torch
from kornia.feature import LoFTR

matcher = LoFTR(pretrained="outdoor")
img0 = torch.rand(1, 1, 480, 640)  # grayscale
img1 = torch.rand(1, 1, 480, 640)

input_dict = {"image0": img0, "image1": img1}
with torch.inference_mode():
    result = matcher(input_dict)

keypoints0 = result["keypoints0"]  # (N, 2)
keypoints1 = result["keypoints1"]  # (N, 2)
confidence = result["confidence"]  # (N,)

Using with ImageStitcher

from kornia.feature import LoFTR
from kornia.contrib import ImageStitcher

matcher = LoFTR(pretrained="outdoor")
stitcher = ImageStitcher(matcher)

Related Pages

Implements Principle

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment