---
title: 'ALIKED Detector: Sparse Keypoint Extraction'
url: https://www.emergentmind.com/topics/aliked-detector
type: topic
---

# ALIKED Detector: Sparse Keypoint Extraction

The ALIKED detector is a deep neural architecture for joint extraction of image keypoints and descriptors, designed to provide computationally efficient and geometrically invariant feature representations for visual measurement tasks. By leveraging a Sparse Deformable Descriptor Head (SDDH) and relaxing traditional dense loss formulations to a sparse regime, ALIKED advances the state of the art in image matching, 3D reconstruction, and visual relocalization, explicitly focusing on efficiency while maintaining high expressiveness.

## 1. Architectural Overview and Motivation

ALIKED aims to resolve the inefficiencies of dense descriptor extraction, which are prevalent in conventional methods such as SIFT, SuperPoint, and dense CNN-based approaches. The detector fuses a Differentiable Keypoint Detection (DKD) module with the SDDH, confining descriptor processing to detected keypoints rather than entire image grids.

Key design objectives include:
- Efficient extraction of keypoints and descriptors suitable for real-time and resource-constrained applications, such as SLAM and mobile visual localization.
- Enabling geometric invariance in descriptors by adaptively sampling supporting features with learned, deformable offsets.

This focus on sparse extraction contrasts with older methods that generate dense score maps and descriptor maps for each pixel, allowing resource-efficient computation without sacrificing performance.

## 2. Sparse Deformable Descriptor Head (SDDH)

The SDDH is a fundamental innovation of ALIKED, designed to model freeform geometric changes through learned sampling positions. For each detected keypoint, descriptor extraction is performed on a local patch (e.g., 5×5 pixels) using the following procedure:

- M deformable sample positions are predicted via a neural subnetwork:
  $$
  p^s = \text{conv}_{1\times 1}\big(\text{SELU}(\text{conv}_{K\times K}(F_{K\times K}))\big)
  $$
  where $p^s \in \mathbb{R}^{M \times 2}$ encodes the offsets for bilateral sampling.
- Bilinear sampling is conducted on the global feature map at these offsets, followed by encoding with $\Phi(\cdot)$ and aggregation using learned weights $w_M$:
  $$
  d = \sum_{i=1}^{M} w_M(p_i) \cdot \Phi(F(p + p_i^s))
  $$
- The number of locations M is tunable, decoupled from the fixed $K\times K$ convolution grid, which enhances the flexibility to model non-affine deformations.

This patch-sparse, position-adaptive descriptor formulation increases the expressiveness of descriptors while minimizing redundant computation.

## 3. Descriptor Extraction Mechanism

ALIKED extracts robust descriptors only at sparse, salient keypoint locations, diverging from the dense sampling paradigm. This enables:

- Computational savings, as the network does not process all pixel positions but only those proximal to keypoints.
- Increased network capacity per descriptor, facilitating stronger geometric invariance and discriminative power.
- High frame rates and lower GFLOPs consumption, benchmarking at over 125 FPS for 640×480 images with 1,000 keypoints for the ALIKED-T(16) variant.

This approach enables a targeted, high-throughput feature pipeline with minimal loss in precision.

## 4. Sparse Neural Reprojection Error (NRE) Loss

Traditional dense loss formulations, such as the Neural Reprojection Error (NRE), require matching likelihoods for all pixels, impeding efficiency. ALIKED introduces a sparse adaptation:

- For a keypoint descriptor $d_A$ in image A and a set of descriptors $D_B$ in image B, compute the similarity vector:
  $$
  \text{sim}(d_A, D_B) = D_B \cdot d_A
  $$
- Compute the matching probability vector via softmax with temperature $t_{des}$:
  $$
  q_m(d_A, D_B) = \text{softmax}\left(\frac{\text{sim}(d_A, D_B) - 1}{t_{des}}\right)
  $$
- The sparse NRE loss for a keypoint $p_A$ is:
  $$
  \mathcal{L}_{ds}(p_A, I_B) = -\ln\left(q_m(d_A, d_B)\right)
  $$
  where $d_B$ is the descriptor at the matched keypoint in image B.

This relaxation to sparse probability vectors reduces memory demand and focuses optimization on keypoint-centric representations.

## 5. Quantitative Performance Evaluation

ALIKED exhibits competitive or superior results across visual measurement benchmarks:

| Task/Benchmark                   | Model/Variant        | Metric(s)         | Performance               |
|--------------------------|---------------------|---------------------|---------------------------|
| Hpatches (Homography)     | ALIKED-T(16)        | MHA@3               | 78.70%                    |
| IMW (3D Reconstruction)   | ALIKED-T/N variants | MS, repeatability   | Higher than DISK, ASLFeat |
| Aachen Day-Night (Reloc.) | ALIKED-N(32)        | Accuracy at 0.25m, 2° | Nearly 100% correct matches |

High frame-rate (>125 FPS) and low computational overhead (GFLOPs) further substantiate its suitability for real-time applications.

## 6. Mathematical Formulation of Core Operations

The ALIKED detector is grounded in several key mathematical formulations:

- Deformable Transformation:
  $$
  [x', y']^T = [x, y]^T + [\Delta x, \Delta y]^T
  $$
- DCN-style deformable convolution (for context):
  $$
  F'(p) = \sum_{i=1}^{K^2} w(p_i) \cdot F(p + p_i + \Delta p_i)
  $$
- Learned deformable position offsets, encoding function, and aggregation (see Section 2 above).
- Sparse NRE loss and softmax-based similarity metrics (see Section 4 above).

These formulations delineate both the feature extraction and the loss-driven optimization processes central to ALIKED.

## 7. Outlook and Future Development

The paper proposes several future directions:

- Integration of joint training for keypoint detection and descriptor extraction in a unified pipeline.
- Optimization for deployment on embedded or mobile platforms, potentially requiring quantization of descriptors.
- Enhancement of deformable descriptor extraction to handle greater variances in viewpoint and scale, suggesting multi-layer or deeper sampling architectures.

These advances target further improvements in geometric invariance, memory efficiency, and hardware compatibility, with application scope extending to all tasks demanding robust, real-time feature extraction.

---

In summary, ALIKED introduces sparse keypoint-based descriptor extraction via deformable sampling, relaxation of dense loss formulations, and comprehensive efficiency benchmarks, establishing a foundation for high-performance, resource-efficient visual measurement pipelines [2304.03608].

Source: https://www.emergentmind.com/topics/aliked-detector