Research map / Detection and segmentation / CNN features

RDD: Robust Feature Detector and Descriptor using Deformable Transformer

CVPR 2025 · Detection and segmentation › CNN features · approach: Context for hard detection conditions · 8 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RDD, a two-branch keypoint detector and descriptor (CNN keypoint branch, deformable-transformer descriptor branch) with an epipolar-based semi-dense refinement, trained on MegaDepth and a new Air-to-Ground dataset.

Problem

Robust keypoint detection and description under large viewpoint, scale, and illumination changes, where local convolutional features fail to capture long-range geometric cues.

Approach

Decouple keypoint detection (CNN with DKD) from description (ResNet-50 plus deformable transformer encoder), so the descriptor branch uses deformable attention to model global context and geometric invariance efficiently, and refine coarse matches to semi-dense using epipolar geometry from sparse matches.

Key results

• On MegaDepth-View, RDD with MNN reaches 38.3 AUC@5 versus 33.1 for DeDoDe-V2-G, and on the Air-to-Ground benchmark it reaches 41.4 AUC@5 versus 31.5 for DeDoDe-V2-G.
• On Aachen Day-Night night split, RDD reaches 86.7 / 92.9 / 99.0 versus 85.7 / 91.8 / 96.9 for ALIKED.
• Ablation: joint training of both branches drops MegaDepth-1500 AUC@5 from 48.2 to 42.8; removing the keypoint branch drops it to 44.1.

Measured effects reported in the paper

Surprising findings

• Joint training of the descriptor and keypoint branches significantly hurts performance, supporting separate training.
• RDD with the dense matcher setting (RDD*) underperforms dense methods such as RoMa and DKM on MegaDepth, and the paper notes semi-dense refinement can fail without confident sparse priors.

Limitations

• No explicit data augmentation during training.
• Semi-dense matching depends on confident sparse correspondences and may fail when they are poor.
• Air-to-Ground dataset collection and usage is only described briefly in the main text, with details deferred to supplementary material.

Experimental setup

Datasets: MegaDepth Air-to-Ground (collected) HPatches Aachen Day-Night

Benchmarks: MegaDepth-1500 MegaDepth-View Air-to-Ground HPatches Aachen Day-Night

Metrics: AUC@5/10/20 pose error mean homography accuracy AUC of localization poses

Baselines: SuperPoint DISK ALIKED XFeat DeDoDe-V2-G DKM RoMa LoFTR ASpanFormer ELoFTR SuperGlue LightGlue

Builds on: DKD (differentiable keypoint detection) deformable attention (Deformable DETR) ALIKED DeDoDe LightGlue

Closest papers by meaning

• Enhancing Deformable Local Features by Jointly Learning To Detect and Describe Keypoints (CVPR 2023 · similarity 0.86)
• Deformable DETR: Deformable Transformers for End-to-End Object Detection (ICLR 2021 · similarity 0.86)
• Extracting Deformation-Aware Local Features by Learning to Deform (NeurIPS 2021 · similarity 0.86)
• ASLFeat: Learning Local Features of Accurate Shape and Localization (CVPR 2020 · similarity 0.85)
• DV-Matcher: Deformation-based Non-rigid Point Cloud Matching Guided by Pre-trained Visual Features (CVPR 2025 · similarity 0.83)
• Key-Grid: Unsupervised 3D Keypoints Detection using Grid Heatmap Features (NeurIPS 2024 · similarity 0.83)