RetailGlue: Semantic Product-Level Image Stitching for Retail Shelf Panoramas

88.4%

F1 Score (ViT-B/16)

20×

Speedup vs Dense

7×

Speedup vs Sparse

32,958

Annotated Correspondences

919

Image Pairs

Abstract

Panoramic shelf images are fundamental to retail analytics, yet constructing them remains a persistent challenge due to repetitive product patterns that confound stitching methods based on local feature matches. We propose a product-level image stitching pipeline that replaces ambiguous low-level feature matching with semantic product correspondences. A YOLOv11l detector trained on the SKU-110K dataset localizes products on individual frames, while DINOv3 embeddings serve as semantic descriptors for a fine-tuned LightGlue matcher. This detect-then-match approach elegantly avoids the detection quality degradation inherent to local feature matching pipelines. Pairwise correspondences are seamlessly extended to full panorama construction via a graph-based multi-image stitching framework that naturally partitions discontinuous sequences to maintain geometric integrity. We evaluate three DINOv3 variants, demonstrating that our semantic approach consistently outperforms baselines relying on local feature matching, with the ViT-B/16 variant achieving an end-to-end stitching F1 score of 88.4%. Furthermore, by drastically reducing the keypoint search space to semantic product instances, our pipeline achieves a remarkable 20× overall runtime speedup compared to state-of-the-art dense matchers, and a 7× speedup over sparse baselines.

Method Overview

RetailGlue inverts the conventional detect-after-stitch paradigm. Rather than applying detection to large, potentially distorted panoramas, we detect products on individual frames first and let those semantic identities drive the matching process.

Figure 1. Comparison of image stitching pipelines. The local feature matching pipeline utilizes low-level keypoints and descriptors, whereas RetailGlue employs semantic product identities for robust and efficient alignment.

Product Detection

YOLOv11l trained on SKU-110K localizes products on individual frames.

Embedding Extraction

DINOv3 [CLS] tokens encode appearance invariant to viewpoint & lighting.

Product-Guided Matching

Fine-tuned LightGlue establishes correspondences using semantic descriptors.

Graph Construction

Pairwise matches build a weighted graph; redundant bridges are pruned.

Panorama Composition

Connected components yield geometrically consistent panoramas via BFS warping.

Contributions

RetailGlue A novel product-level LightGlue adaptation using DINOv3 semantic embeddings as descriptors, specifically designed for robust retail image stitching.
Pipeline An inverted detect-then-match pipeline that preserves object detection quality by operating on individual frames rather than large stitched panoramas.
Graph A graph-based multi-image stitching framework with robustness to discontinuous captures via graph partitioning, achieving up to a 20× runtime speedup.
Dataset RemRetail100_v2: an extended benchmark with 32,958 annotated product correspondences across 919 image pairs.

Results

Product Matching Quality

We evaluate LightGlue fine-tuned with different DINOv3 variants against a DINOv2 baseline. DINOv3 ViT-L/16 achieves the best recall (79.4%), critical for reliable geometric alignment.

DINO Variant	Desc. Dim.	Precision (%)	Recall (%)	F1 (%)
DINOv2 ViT-S/14	384	69.7	61.5	65.3
DINOv3 ViT-S/16	384	78.2	71.9	74.9
DINOv3 ViT-B/16	768	79.1	68.4	73.4
DINOv3 ViT-L/16	1024	76.8	79.4	78.1

Table 1. Product matching performance for fine-tuned LightGlue models. Bold indicates best in column.

End-to-End Stitching Quality on RemRetail100_v2

RetailGlue with DINOv3 ViT-B/16 achieves the best end-to-end F1 of 88.4%, surpassing all local feature matching baselines including the dense matcher RoMav2.

Pipeline / Matcher	Dim.	Acc.	Prec.	Rec.	F1
Local feature matching pipelines (baselines)
SuperGlue	256	78.7	89.9	85.3	86.5
SuperPoint + LightGlue	256	80.2	89.5	86.9	87.8
DISK + LightGlue	128	73.9	85.9	82.5	83.3
MINIMA + LightGlue	256	79.8	90.1	86.6	87.6
RoMav2	—	80.3	90.4	86.9	88.1
SuperPoint + GlueStick	256	79.5	89.6	86.5	87.4
SuperPoint + LightGlueStick	256	80.2	90.2	87.0	87.9
RetailGlue (Ours)
DINOv2 ViT-S/14 + LightGlue	384	79.9	88.3	87.6	87.9
DINOv3 ViT-S/16 + LightGlue	384	79.9	89.4	87.4	88.1
DINOv3 ViT-B/16 + LightGlue	768	80.7	89.4	87.7	88.4
DINOv3 ViT-L/16 + LightGlue	1024	80.2	89.7	87.4	88.3

Table 2. End-to-end stitching quality on RemRetail100_v2. All RetailGlue variants use the SKU-110K YOLO detector.

Runtime Comparison

Product localization and descriptor extraction are inherent prerequisites for retail analytics and add no extra overhead. The graph construction phase is the true bottleneck — and RetailGlue's semantic sparsity makes it negligible.

Sequence	Graph Const. (s)	Refinement (s)	Composition (s)	Total (s)
Short (4 images)
RoMav2	3.58	0.01	0.45	4.03
DISK + LightGlue	4.85	0.03	0.68	5.56
RetailGlue (ours)	0.48	0.01	0.67	1.16
Medium (8 images)
RoMav2	15.12	0.01	0.69	15.82
DISK + LightGlue	9.11	0.05	0.60	9.76
RetailGlue (ours)	1.58	0.03	0.73	2.34
Long (13 images)
RoMav2	59.63	0.62	0.18	60.43
DISK + LightGlue	20.57	0.59	0.20	21.36
RetailGlue (ours)	2.34	0.51	0.18	3.03

Table 3. Runtime comparison (seconds). RetailGlue achieves 7× speedup over DISK+LightGlue and 20× over RoMav2 on long sequences.

Qualitative Results

Side-by-side comparisons of DISK + LightGlue (baseline) and our RetailGlue pipeline. Use the arrows to browse examples.

✓ Success Cases

DISK + LightGlue Success 1 — DISK + LightGlue

RetailGlue (Ours) Success 1 — RetailGlue

DISK + LightGlue Success 2 — DISK + LightGlue

RetailGlue (Ours) Success 2 — RetailGlue

DISK + LightGlue Success 3 — DISK + LightGlue

RetailGlue (Ours) Success 3 — RetailGlue

DISK + LightGlue Success 4 — DISK + LightGlue

RetailGlue (Ours) Success 4 — RetailGlue

DISK + LightGlue Success 5 — DISK + LightGlue

RetailGlue (Ours) Success 5 — RetailGlue

Success cases. In densely packed, highly repetitive retail scenes, the local feature matcher fails due to structural ambiguity, while our semantic product-level approach resolves identical products for a perfectly aligned panorama.

✗ Failure Cases

DISK + LightGlue Failure 1 — DISK + LightGlue

RetailGlue (Ours) Failure 1 — RetailGlue

DISK + LightGlue Failure 2 — DISK + LightGlue

RetailGlue (Ours) Failure 2 — RetailGlue

DISK + LightGlue Failure 3 — DISK + LightGlue

RetailGlue (Ours) Failure 3 — RetailGlue

DISK + LightGlue Failure 4 — DISK + LightGlue

RetailGlue (Ours) Failure 4 — RetailGlue

Failure cases. Scenes with large-scale products and narrow overlap, where our product-level approach lacks sufficient bounding boxes for a reliable homography, while the baseline leverages abundant textural details on large packaging.

Figure 2. Multi-panorama generation via graph partitioning. Given a challenging 8-frame sequence with insufficient overlap in the middle frames, our graph-based pipeline naturally divides the sequence into three geometrically consistent sub-panoramas rather than forcing a distorted global stitch.

RemRetail100_v2 Dataset

We introduce an extended retail benchmark specifically designed for image stitching evaluation. All data is publicly released on the Hugging Face Hub.

253

Sequences

1,172

Total Images

919

Image Pairs

32,958

Correspondences

35.9

Avg Matches/Pair

Ground truth correspondences were established through a semi-automatic process: initial matches were generated via brute-force matching, then manually verified and corrected using a custom annotation interface. The dataset is split 70/20/10 at the sequence level to prevent data leakage.

Downloads

The benchmark release is organized into four complementary dataset repositories on Hugging Face:

Dataset	Contents	Link
RemRetail100_v2	Raw shelf images for end-to-end stitching evaluation	RemRetail100_v2
LightGlue · DINOv3 ViT-S/16	Images + 384-d embeddings + pair matches	RetailGlue-lightglue-dinov3-vits
LightGlue · DINOv3 ViT-B/16	Images + 768-d embeddings + pair matches	RetailGlue-lightglue-dinov3-vitb
LightGlue · DINOv3 ViT-L/16	Images + 1024-d embeddings + pair matches	RetailGlue-lightglue-dinov3-vitl

Figure 5. Example image pairs from RemRetail100_v2. Visualizations display overlapping shelf scenes alongside their manually verified, ground truth product matches.

BibTeX

If you find our work useful, please cite:

@inproceedings{oztuner2026retailglue,
  title     = {RetailGlue: Semantic Product-Level Image Stitching
               for Retail Shelf Panoramas},
  author    = {\"Ozt\"uner, Arda and \c{C}elik, Ayberk and
               Yal\c{c}{\i}ner, \.{I}brahim \c{S}amil and Calap, Server},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision
               and Pattern Recognition Workshops (CVPRW) -- Image Matching:
               Local Features {\&} Beyond},
  year      = {2026}
}

Acknowledgements

We thank the open-source communities behind YOLOv11, DINOv3, and LightGlue. We also gratefully acknowledge the RemPeople family for their valuable support and contributions to this work, with special thanks to Rem People.