Teaching DINOv3 About Partial 3D Geometry: A Self-Supervised Geometry-Aware Approach

Viktoria Ehm; Dongliang Cao; Riccardo Marin; Daniel Scholz; Weikang Wang; Florian Bernard; Daniel Cremers

cs.CVOct 15, 2026

Teaching DINOv3 About Partial 3D Geometry: A Self-Supervised Geometry-Aware Approach

Authors: Viktoria Ehm, Dongliang Cao, Riccardo Marin, Daniel Scholz, Weikang Wang, Florian Bernard, Daniel Cremers

Organizations: Technical University of Munich · Munich Center for Machine Learning · University of Bonn

Abstract

Partial shape matching is a crucial yet underexplored problem in 3D vision, with significant relevance to real-world scenarios where shapes are often only partially observed. Existing feature descriptors face difficulties in this setting, as traditional representations either struggle with the boundaries of partial shapes or heavily depend on the shape's spatial position. While existing approaches have employed DINO features for partial shape matching, these features are not inherently suited for handling partial observations. In this work, we propose a method to refine DINO features using LoRA-based self-supervised learning, enabling the generation of feature descriptors that are robust to partiality. Our features substantially improve performance on partial shape matching compared to traditional or vision foundation features. Additionally, when integrated into existing partial shape matching pipelines, we achieve state-of-the-art results on partial shape matching and left-right prediction benchmarks.

Figures & tables

Figure 2: Our Self-Supervised Training Pipeline: Given a full mesh $\mathcal{X}$ , we first generate a partial observation $\mathcal{Y}$ . After randomly rotating both shapes with different rotation matrices around the Y-axis, we render the partial and full shape from different perspectives with cameras $\mathcal{C}$ . For every camera $c^{i}\in\mathcal{C}$ , we obtain an image $\mathbf{I}^{i}_{\mathcal{X}}$ of the full shape, which is then passed through a frozen DINOv3 architecture with a LoRA module with rank $r$ to extract pixel-wise features $\mathbf{Q}^{i}_{\mathcal{X}}$ . These are then backprojected on the shape and averaged to get vertex-wise features $\mathbf{F}_{\mathcal{X}}$ . Similarly, we proceed for the partial shape $\mathcal{Y}$ . Finally, we use the Geodesic-Aware PointInfoNCE contrastive loss to align corresponding vertices of both shapes and separate all others based on geodesic distance. In our experiments we use $4$ views for the partial and $9$ views for the full shape.

Type	Dataset	WKS	HKS	Diff3f	DINOv2	DINOv3	Trained on	GeoLoRA
P2F	BeCoS [ 23 ]	58.35	69.50	25.61	21.54	19.95	BeCoS	5.57
	Shrec16 Cuts [ 16 ]	62.08	53.32	28.01	26.58	21.25	BeCoS *	10.86
	Shrec16 Holes [ 16 ]	50.33	65.56	26.67	20.22	20.22	BeCoS *	9.70
	Pfaust-M [ 8 ]	50.28	52.19	12.60	11.38	11.36	Faust	2.88
	Pfaust-H [ 8 ]	54.75	60.83	19.32	17.47	13.82	Faust	3.33
P2P	BeCoS [ 23 ]	41.35	40.48	19.45	17.43	16.84	BeCoS	6.72

Table 1: Improvement of Feature Quality Through Our Self-Supervised Pipeline: We show that with GeoLoRA we can improve the feature quality on all partial datasets in terms of the mean geodesic error (

\times 100

). BeCoS * indicates BeCoS without Shrec16 data.

Panel of Figure 3 — Figure 3: Qualitative Results of Feature Quality: We show colour-coded correspondences based on the features computed with different feature extractors. GeoLoRA exhibits improved performance, especially in left-right predictions.

Geo Err ( $\downarrow$ )	Features	ULRSSM [ 12 ]	DPFM [ 3 ]
BeCoS	DINOv3	26.97 (26.19)	23.16
BeCoS	GeoLoRA	11.62 (11.54)	13.38
Shrec16Cuts	DINOv3	5.94 (4.30)	10.78
Shrec16Cuts	GeoLoRA	3.01 (1.97)	6.57
Shrec16Holes	DINOv3	12.41 (9.66)	10.41
Shrec16Holes	GeoLoRA	8.07 (6.88)	7.38

Table 2: Partial-to-Full Shape Matching with GeoLoRA: We show that we can improve the downstream task performance of partial-to-full shape matching with GeoLoRA in comparison to DINOv3 features in terms of mean geodesic error (

\times 100

). The numbers in brackets of ULRSSM show the geodesic error after test-time adaptation.

	Features	EchoMatch [ 62 ]		DPFM [ 3 ]
		Err	mIoU	Err	mIoU
BeCoS	DINOv3	9.74	67.07	9.83	49.80
BeCoS	GeoLoRA	5.55	71.27	8.20	56.04
Psmal	DINOv3	5.56	82.75	11.25	73.25
Psmal	GeoLoRA	4.33	85.41	8.91	78.95
CP2P24	DINOv3	2.97	87.02	3.89	71.55

Table 3: Partial-to-Partial Shape Matching with GeoLoRA: We demonstrate that our GeoLoRA features achieve better downstream partial-to-partial shape matching performance than DINOv3, as measured by geodesic error (

\times 100

) and Intersection over Union (IoU).

Panel of Figure 4 — Figure 4: Qualitative Results of Partial-to-Full Shape Matching Methods: We show one qualitative result of partial-to-full shape matching on the Shrec16 Cuts dataset. We can see improvement in the quality of the matching, especially at the arm with GeoLoRA features.

	16	25	36	49	64	100
Geo Err	25.47	20.80	21.48	20.09	20.29	20.00
Time(s)	1.52	2.35	3.36	4.59	6.01	9.47

Table 4: Number of Views

N

with Frozen DINOv3: With an increasing number of views, the time increases linearly, while the geodesic error (

\times 100

) already saturates earlier. Thus, we choose

49

views for our experiments.

	PointInfoNCE	GeoLoRA
Err ( $\downarrow$ )	11.29	5.89

Table 5: PointInfoNCE vs GeoLoRA: We show that our loss formulation decreases the geodesic error (

\times 100

) significantly compared to the widely used PointInfoNCE loss.

	DINOv3	PointInfoNCE	GeoLoRA
Acc ( $\uparrow$ )	70.45	84.09	91.42

Table 6: Chirality Score for BeCoS Partial Shapes: We show that we can improve the left-right prediction of [ 58 ] with GeoLoRA in terms of accuracy.

Panel of Figure 5 — Figure 5: Qualitative Results of Left-Right Prediction: GeoLoRA shows better chirality prediction than DINOv3 (shapes from BeCoS ).

Panel of Figure 6 — Figure 6: Real-World Scan: We match a real-world partial scan on a template human. DINOv3 struggles to characterise incomplete geometry, while GeoLoRA shows robust performance.

Symbol	Description
$\mathcal{X}=(\mathbf{V}_{\mathcal{X}},\mathbf{T}_{\mathcal{X}})$	3D Full Shape
$\mathbf{V}_{\mathcal{X}}$	Vertices of shape $\mathcal{X}$
$\mathbf{T}_{\mathcal{X}}$	Triangles of shape $\mathcal{X}$
$\mathcal{Y}=(\mathbf{V}_{\mathcal{Y}},\mathbf{T}_{\mathcal{Y}})$	3D Partial Shape, $\mathbf{V}_{\mathcal{Y}}\subset\mathbf{V}_{\mathcal{X}}$ , $\mathbf{T}_{\mathcal{Y}}\subset\mathbf{T}_{\mathcal{X}}$
$\mathbf{\Pi}_{\mathcal{Y}\mathcal{X}}$	Correspondence from partial to full shape
$\mathcal{C}=(c^{1},...,c^{N})$	Set of all cameras

Table 7: Overview of Symbols

$r$ / $\alpha$	4/8	8/16	16/32
Geo Error ( $\downarrow$ )	8.14	6.56	5.89

Table 8: LoRA Hyperparameter Ablation: We observe the best results with rank

r=16

and

\alpha=32

	DINOv2	DINOv3	Diff3f	GeoLoRA
Time [s] ( $\downarrow$ )	4.72	4.45	360.45	4.59

Table 9: Inference Time Comparison: While DINOv2, DINOv3 and GeoLoRA have similar runtime during inference, Diff3f’s runtime is significantly increased due to the need to additionally run inference with Stable Diffusion.

PFAUST	DINOv3-B	GeoLoRA-B	DINOv3-L	GeoLoRA-L
-M	11.36	2.88	11.40	2.43
-H	13.82	3.33	12.50	2.65

Table 10: Performance of Different Backbone Scales: We show that with a larger backbone we do not gain significant improvement in terms of geodesic error.

Geo Error $(\downarrow)$	XYZ	Ours
BeCoS	40.46 (40.18)	11.62 (11.54)
Shrec16Cuts	7.04 (3.2)	3.01 (1.97)
Shrec16 Holes	13.75 (8.2)	8.07 (6.88)
Pfaust-M	7.65 (7.49)	1.61 (1.61)
Pfaust-H	9.39 (9.24)	2.29 (2.19)

Table 11: Comparison of XYZ features vs GeoLoRA features: We compare the geodesic error

(\times 100)

on different partial-to-full datasets with ULRSSM. On all datasets, our features outperform the originally used XYZ features. The numbers in brackets “

()

” indicate geodesic error results after test-time adaptation [ 12 ] .

	Original	Ours
Chirality Accuracy ( $\uparrow$ )	57.40	91.42

Table 12: Comparison to DINOv2 + Stable Diffusion Features: We compare the chirality (left-right) accuracy of the features used in [ 58 ] to our GeoLoRA features. Our method’s features show higher accuracy.

Panel of Figure 7 — Figure 7: Failure Cases: We show two failure cases of our method. The method primarily struggles with shapes in challenging poses, where the shapes are not positioned upright.

Panel of Figure 8 — Figure 8: Additional Results on Real-World Statues from [ 1 ] : DINOv3 fails to find good correspondences between the partial statues and the full template shape, where GeoLoRA finds reasonably dense correspondences.

Panel of Figure 9 — Figure 9: 2D Keypoint Matching: We show examples for keypoint matching on the SPair-71k [ 42 ] dataset. Our method shows superior matching results in these examples.

Panel of Figure 10 — Figure 10: 3D Keypoint Matching: Our method can be used for 3D keypoint matching. We show examples of the BeCoS test set.

Panel of Figure 11 — Figure 11: Qualitative Results of Feature Quality: We show colour-coded correspondences based on the features computed with different feature extractors. GeoLoRA exhibits improved performance, especially in left-right and front-back predictions.

Panel of Figure 12 — Figure 12: More Qualitative Results: We show more qualitative results of our partial-to-full features.

Panel of Figure 13 — Figure 13: Qualitative Results of Partial-to-Partial Matching: We show colour-coded correspondences for partial-to-partial shape matching methods. We compare DINOv3 features with GeoLoRA features. The red area indicates points that are predicted not to be in the overlapping region. On the BeCoS dataset and the Psmal dataset, we demonstrate improved performance with GeoLoRA features compared to DINOv3 features. On CP2P24 we observe similar matching quality.

Panel of Figure 14 — Figure 14: Qualitative Results of Partial-to-Full Shape Matching Methods: We show two more examples of partial-to-full shape matching on the Pfaust-H and Shrec16 Holes dataset. We can see improved matching quality with our GeoLoRA features, especially in the extremities of the shapes, such as the left arm of the human and the legs of the horse.

3D Representation Learning

3D representation learning builds geometric features for recognizing, reconstructing, generating, and reasoning about objects and scenes.

See all 191 papers

TokenMatch: 3D Mesh Correspondence Transformer with Curvature-Guided Tokenisation
Sep 3, 2026Adeela Islam, Zorah Lähner, Vittorio Murino +13D ViTs Non-Rigid Shape Correspondence
1Italian Institute of Technology · University of Genoa · 6Max Planck Institute for Informatics, SIC +4
DinoComplete: 3D Shape Completion with Distilled Semantic Priors and State Space Models
May 26, 2026Furkan Mert Algan, Eckehard Steinbach3D Reconstruction 3D Shape Completion
Chair of Media Technology · Munich Institute of Robotics and Machine Intelligence · School of Computation Information and Technology, Technical University of Munich
Self-Supervised Learning of Semantically Consistent Visual Features via a Simple 3D Prior
Oct 12, 2026Sebastian Day, Oisin Mac AodhaSelf-Supervised Visual Representation Learning Semantic Correspondence
University of Edinburgh

Sep 3, 2026J/K move · Enter open · S save

TokenMatch: 3D Mesh Correspondence Transformer with Curvature-Guided Tokenisation

Adeela Islam, Zorah Lähner, Vittorio Murino, Vladislav Golyanik

1Italian Institute of Technology · University of Genoa · 6Max Planck Institute for Informatics, SIC+4

3D ViTs Non-Rigid Shape Correspondence 3D Shape Matching

Open paper AskPDF arXiv

While data-driven 3D shape correspondence estimation has recently seen substantial progress, robust matching under partial observations and strong non-isometric deformations remains challenging. Existing learning-based approaches often rely on hand-crafted descriptors or template-based representations, whereas recent generative models over functional maps suffer from high inference cost, limited interpretability, and poor generalisation to partial shapes. In response to these limitations, this paper introduces TokenMatch, a new transformer-based unified model for estimating 3D shape correspondences. Our feed-forward approach trained exclusively on BeCoS, a challenging non-isometric partial-to-partial shape-matching dataset, can generalise to matching full shapes without retraining or fine-tuning. TokenMatch uses self- and cross-attention mechanisms to efficiently learn patch-level and point-level relations as well as dense correspondences between shape pairs. Our core insight is that meshes can be adaptively tokenised into patches using shape curvature guidance, enabling effective learning of shape-specific geometric descriptors for correspondence estimation. We evaluate TokenMatch on standard benchmarks for partial and full shape matching, including CP2P, PSMAL, BeCoS, FAUST, SCAPE, and SHREC'19. Our method achieves consistently high performance, in most cases outperforming existing methods for partial and full shape matching in the mean geodesic error and intersection-over-union metrics, while also running faster at sub-second inference speeds.

Abstract

3D Representation Learning

Explore similar work

TokenMatch: 3D Mesh Correspondence Transformer with Curvature-Guided Tokenisation

DinoComplete: 3D Shape Completion with Distilled Semantic Priors and State Space Models

Self-Supervised Learning of Semantically Consistent Visual Features via a Simple 3D Prior