SRMatcher: Semantic-Aware Matching

• SRMatcher is a detector-free feature matching framework that integrates semantic cues from large vision foundation models to improve homography estimation.
• It employs a two-stage semantic-aware fusion strategy, combining intra- and cross-image interactions, to enhance feature representation and reduce false correspondences.
• Experimental evaluations demonstrate that SRMatcher outperforms traditional methods like LoFTR and GeoFormer, achieving higher AUC and better corner accuracy across benchmarks.

SRMatcher is a detector-free feature matching framework for homography estimation that explicitly integrates fine-grained semantic information from vision foundation models (VFMs) into the feature representation pipeline. Designed to address the limitations of conventional detector-based and detector-free methods—specifically their lack of semantic awareness—SRMatcher adopts a semantic-aware fusion strategy to achieve high-precision and robust image correspondence, particularly in challenging settings where visual ambiguity and semantic inconsistency can degrade performance. Its architecture is agnostic to the specific transformer matching backbone and can function as a general plug-and-play module, yielding consistent accuracy gains across a variety of dense correspondence tasks (Liu et al., 2024).

1. Homography Estimation Problem Setting

Homography estimation entails recovering a $3\times3$ matrix $H$ (defined up to scale), such that points $x=(u,v,1)^\top$ in image $I_0$ correspond to $x'\sim Hx$ in image $I_1$. This operation is fundamental to tasks such as panorama stitching, SLAM, and augmented reality. Standard performance metrics include:

• Corner Error: Measures the Euclidean distance between projected corners and their ground-truth locations post-homography.
• Area Under the Cumulative error Curve (AUC): Reports the cumulative distribution of corner errors at multiple pixel thresholds (e.g., 1px, 3px, 5px, 10px).

SRMatcher is evaluated under these criteria, with particular focus on corner alignment accuracy at strict thresholds.

2. Detector-Free Matching Paradigm

Traditional feature matching pipelines are detector-based, relying on sparse interest point detection, descriptor extraction, and nearest-neighbor matching, followed by model fitting (e.g., via RANSAC). Their effectiveness diminishes in low-texture or blurred regions due to limited keypoint repeatability. In contrast, detector-free methods (e.g., LoFTR, GeoFormer) operate on dense feature maps, employing self- and cross-attention to generate pixel-level confidence matrices. Pixel correspondences are then extracted via thresholding and mutual nearest neighbor (MNN) filtering. This paradigm is robust to texture paucity and is fully differentiable.

SRMatcher situates itself in the detector-free regime but augments the process with strong semantic priors, transcending limitations of purely geometric or low-level feature matching.

3. Semantic-Aware Feature Representation and Fusion

Semantic inconsistency—matching points that are geometrically similar but semantically unrelated—remains a critical source of false correspondences. SRMatcher addresses this challenge by fusing dense appearance features from the matching backbone with semantic features extracted from a frozen, pre-trained DINOv2 (ViT-B/14) VFM. The fusion is orchestrated via the Cross-image Semantic-aware Fusion Block (SFB), consisting of two Semantic-Guide Interactions Blocks (SGIB):

• Stage 1 (intra-image fusion): Semantic activation $Q_i = \mathrm{SA}(S_i)$ is applied to local features $\hat{C}_i$, followed by a softmax-weighted combination, producing an enriched representation $\tilde{C}_i'$.
• Stage 2 (cross-image fusion): The semantic activation $Q'_j = \mathrm{SA}(S_j)$ from the opposite image drives a further fusion, producing $\tilde{C}_i$, which embeds both local appearance and cross-view semantic cues.

This dual-stage fusion mitigates semantic mismatch and encodes correspondence-relevant semantic structure at full spatial resolution.

4. SRMatcher Architecture and Training Objectives

The SRMatcher pipeline is summarized as follows:

1. Input image pairs $(I_0, I_1)$ are processed by a CNN backbone (ResNetFPN), resulting in coarse features $(C_0, C_1)$ at $1/8$ scale and fine features $(F_0, F_1)$ at $1/2$ scale.
2. Self- and cross-attention layers yield refined appearance features $(\hat{C}_0, \hat{C}_1)$.
3. Semantic features $(S_0, S_1)$ are extracted from the frozen DINOv2 model.
4. The SFB fuses $(\hat{C}_i, S_0, S_1)$ into unified features $(\tilde{C}_i)$.
5. Coarse correspondences are identified via dual-softmax on the similarity matrix, followed by threshold + MNN, yielding matches $M_c$.
6. For each $(i,j)\in M_c$, a local region-based overlap generates fine-grained matches $M_f$ through full window-to-window similarity computation.
7. Homography $H$ is estimated via RANSAC on $M_f$.

Training proceeds in a self-supervised regime, with $I_1$ derived by randomly warping $I_0$ with a known homography. Losses include:

• Coarse Match Loss: $$L_c = -\frac{1}{|G_c|} \sum_{(i,j)\in G_c} \log P_c(i,j)$$
• Fine Match Loss: Analogously defined as $L_f$, for dense correspondences.
• Total Objective: $L_{\rm total}=L_c+L_f$

5. Experimental Evaluation and Results

SRMatcher demonstrates state-of-the-art performance across multiple benchmarks, particularly in strict alignment regimes:

Dataset	Metric	Baseline	Baseline Result	SRMatcher Variant	Result (%)	Gain
HPatches	AUC@1px	LoFTR	34.2	SRMatcher_LoFTR	38.6	+12.9
HPatches	AUC@1px	GeoFormer	44.3	SRMatcher_GeoFormer	49.2	+11.1
ISC-HE	AUC@3px	GeoFormer	20.1	SRMatcher_GeoFormer	20.5	+3.0
MegaDepth-1500	AUC@5°/10°	GeoFormer	34.2/50.5	SRMatcher_GeoFormer	35.8/52.0	various

SRMatcher consistently outperforms prior approaches such as MESA, CasMTR, and SuperGlue across all pixel thresholds. Relative pose estimation tasks further reveal the model's adaptability beyond strict homography estimation. Qualitative evidence indicates fewer outlier matches and improved corner alignments in composite panoramas (Liu et al., 2024).

6. Ablation Studies and Plug-and-Play Generality

Detailed ablations on HPatches with SRMatcher_GeoFormer demonstrate the criticality of each architectural component:

• Removing the SFB reduces AUC@1px from 49.2 to 45.3.
• Replacing SGIB with simple spatial-attention fusion results in 47.8.
• Omitting stage 2 (cross-image fusion): 48.1.
• Substituting DINOv2 with ResNet-50: 47.2.
• Eliminating overlap-based fine matching: 47.6.

Alternative semantic extractors (ResNet-50, CLIP, DINOv1, SAM) were tested, with DINOv2 providing the best performance. Fusion performed after the attention backbone, rather than prior to FPN, strengthens results further. Increasing the number of SGIB stages enhances mean-AUC but incurs additional computational cost.

Notably, the SFB+semantic extractor module can be inserted into other detector-free matching architectures, such as LoFTR or ASpanFormer, consistently yielding measurable precision improvements, underlining the generality of the SRMatcher approach.

7. Implementation and Training Details

• Training Dataset: Oxford-Paris (Oxford5K + Paris6K), augmented via self-supervised homography warping and photometric augmentation (brightness/contrast, blur, noise).
• Optimization: Adam ($\beta_1=0.9$, $\beta_2=0.999$), learning rate $1{\rm e}{-3}$, 15 epochs, batch size 1; hardware: 8×RTX3090.
• Backbone/Variant Support: Compatible with LoFTR, ASpanFormer, GeoFormer, with only the attention module differing among SRMatcher variants.
• Semantic Extractor: DINOv2 (ViT-B/14), kept frozen throughout.

SRMatcher is thus the first detector-free matcher that deeply integrates fine-grained semantic features from large VFMs into the core feature matching backbone. It sets a new state of the art in homography estimation and enhances transformer-based matching pipelines through semantic-aware fusion and robust, differentiable correspondence mechanisms (Liu et al., 2024).