Implementation:Kornia Kornia ScaleSpaceDetector

Knowledge Sources	Kornia
Domains	Vision, Feature_Detection
Last Updated	2026-02-09 15:00 GMT

Overview

Concrete tool for differentiable multi-scale local feature detection provided by Kornia.

Description

ScaleSpaceDetector is a modular nn.Module that assembles 5 sub-modules: scale pyramid generator, response function (cornerness), soft non-maximum suppression, affine shape estimator, and orientation estimator. It detects keypoints as Local Affine Frames (LAFs) of shape (B, N, 2, 3) encoding position, scale, and orientation. The detector is differentiable, enabling end-to-end training of detection and description.

Usage

Initialize with desired number of features and optional custom sub-modules. Pass grayscale images to detect keypoints. Use with descriptor modules (HardNet, SIFTDescriptor) for matching.

Code Reference

Source Location

Repository: kornia
File: kornia/feature/scale_space_detector.py
Lines: L72-264

Signature

class ScaleSpaceDetector(nn.Module):
    def __init__(
        self,
        num_features: int = 500,
        mr_size: float = 6.0,
        scale_pyr_module: Optional[nn.Module] = None,
        resp_module: Optional[nn.Module] = None,
        nms_module: Optional[nn.Module] = None,
        ori_module: Optional[nn.Module] = None,
        aff_module: Optional[nn.Module] = None,
        minima_are_also_good: bool = False,
        scale_space_response: bool = False
    ) -> None

Forward

def forward(
    self,
    img: torch.Tensor,
    mask: Optional[torch.Tensor] = None
) -> Tuple[torch.Tensor, torch.Tensor]

Import

from kornia.feature import ScaleSpaceDetector

I/O Contract

Inputs

Name	Type	Required	Description
img	torch.Tensor	Yes	Grayscale image tensor of shape (B, 1, H, W)
mask	Optional[torch.Tensor]	No	Valid region mask for restricting detection area

Outputs

Name	Type	Description
lafs	torch.Tensor (B, N, 2, 3)	Local Affine Frames encoding keypoint position, scale, and orientation
responses	torch.Tensor (B, N)	Detector response scores for each keypoint

Usage Examples

Basic Detection

import torch
from kornia.feature import ScaleSpaceDetector

detector = ScaleSpaceDetector(num_features=500)
img = torch.rand(1, 1, 256, 256)

lafs, responses = detector(img)
# lafs.shape == (1, 500, 2, 3)
# responses.shape == (1, 500)

Detection with Custom Descriptor

import torch
from kornia.feature import ScaleSpaceDetector, HardNet, LAFDescriptor

detector = ScaleSpaceDetector(num_features=500)
descriptor = LAFDescriptor(HardNet())

img = torch.rand(1, 1, 256, 256)
lafs, responses = detector(img)
descriptors = descriptor(img, lafs)
# descriptors.shape == (1, 500, 128)

Related Pages

Implements Principle

Page Connections

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

Principle

Implementation

Heuristic

Environment