D2-Net: A Trainable CNN for Joint Description and Detection of Local Features

By Mihai Dusmanu, Ignacio Rocco, Tomás Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, Torsten Sattler. Computer Vision and Pattern Recognition, 2019

924 citations recorded by the source corpus.

In this work we address the problem of finding reliable pixel-level correspondences under difficult imaging conditions. We propose an approach where a single convolutional neural network plays a dual role: It is simultaneously a dense feature descriptor and a feature detector. By postponing the detection to a later stage, the obtained keypoints are more stable than their traditional counterparts based on early detection of low-level structures. We show that this model can be trained using pixel correspondences extracted from readily available large-scale SfM reconstructions, without any further annotations. The proposed method obtains state-of-the-art performance on both the difficult Aachen Day-Night localization dataset and the InLoc indoor localization benchmark, as well as competitive performance on other benchmarks for image matching and 3D reconstruction.

Subfields: Aerospace Engineering, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/0ca14ad73196181f363968c8320aba196865c6e6

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Aerospace Engineering

In Computer Vision and Pattern Recognition

Citations in the corpus

COTR: Correspondence Transformer for Matching Across Images (2021, background): “The accuracy of the depth maps is sufficient to learn accurate local features, as demonstrated by [16, 54, 71].”
Continuous 3D Perception Model with Persistent State (2025, neutral): “…replacing hand-crafted components with learning-based alternatives (e.g., substituting conventional feature descriptors [6, 47] with learned ones [17, 20, 63, 73]), integrating data-driven priors [13, 77–79, 93, 97, 105] into the systems, or optimizing entire systems end-to-end [76, 78, 83, 94].”
A Tale of Two Features: Stable Diffusion Complements DINO for Zero-Shot Semantic Correspondence (2023, background): “Deep models [14, 37, 42, 49, 63, 68] learn effective features for dense correspondence that can be robust to photometric and geometric changes such as rotation, scaling, or perspective transformation.”
Learning Target Candidate Association to Keep Track of What Not to Track (2021, background): “Hence, in the domain of sparse keypoint matching, recent learning based approaches leverage large scale datasets [19, 38] such as MegaDepth [36] or ScanNet [9], that provide ground truth matches.”
Depth Anything 3: Recovering the Visual Space from Any Views (2025, positive): “Early learning methods injected robustness at the component level: learned detectors [20], descriptors for matching [22], and differentiable optimization layers that expose pose/depth updates to gradient flow [31, 33, 62].”
Image Matching from Handcrafted to Deep Features: A Survey (2020, methodology): “D2-Net (Dusmanu et al. 2019) uses a single CNN to play a dual role: simultaneously achieving a dense feature descriptor and a feature detector.”
Patch-NetVLAD: Multi-Scale Fusion of Locally-Global Descriptors for Place Recognition (2021, methodology): “Traditional handcrafted local feature methods such as SIFT [44], SURF [6] and ORB [60], and more recent deep-learned local features like LIFT [84], DeLF [52], SuperPoint [18] and D2Net [20], have been extensively employed for VPR [52, 9, 17], visual SLAM [50] and 6-DoF localization [63, 77, 20, 61].”
Unifying Deep Local and Global Features for Image Search (2020, methodology): “In particular, note that we do not require patch-level supervision to train local features, unlike most recent works [14,49,36,29].”
Grounding Image Matching in 3D with MASt3R (2024, methodology): “To obtain reliable pixel correspondences from pointmaps, a standard solution is to look for reciprocal matches in some invariant feature space [26,78,102, 106].”
ASpanFormer: Detector-Free Image Matching with Adaptive Span Transformer (2022, background): “Differing from detector-based matching methods, which typical involve detecting [11,9,10,8], describing [26,27,28,12,29] and matching [30,31,32,33,34,35,36] a set of keypoints, detector-free matching consumes a pair of images and output correspondences in one shot.”
Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass (2025, neutral): “For example, [14, 19, 25, 42, 51, 68] improve feature extraction and correspondences, [27, 50, 59, 71] learn to estimate camera pose, and [52] introduce a bundle adjustment layer.”
Learning and aggregating deep local descriptors for instance-level recognition (2020, background): “In contrast to other learned local feature detectors [13] that use keypoint-level supervision and non-maxima suppression, DELF features are not precisely localized and suffer from redundancy since deep network activations are typically spatially correlated.”
D3Feat: Joint Learning of Dense Detection and Description of 3D Local Features (2020, methodology): “To this end, we draw inspiration from D2-Net [7] in 2D domain for a joint learning of a feature detector and descriptor.”
DUSt3R: Geometric 3D Vision Made Easy (2023, positive): “These improvements encompass advanced feature description [26, 28, 97, 135, 167], more accurate image matching [3, 17, 60, 82, 100, 120, 126, 145], featuremetric refinement [59], and neural bundle adjustment [58, 153].”
VGGT: Visual Geometry Grounded Transformer (2025, neutral): “In recent years, deep learning has improved many components of the SfM pipeline, with keypoint detection [12, 19, 76, 102] and image matching [7, 46, 60, 66] being two primary areas of focus.”