Establishing dense correspondence between high-resolution 3D shapes remains challenging due to the spatial--spectral coupling at the core of many functional-map methods, which grows quadratically with mesh resolution. We observe that this coupling is equivalent to a scaled dot-product attention formulation, with learned descriptors serving as queries and keys and spectral basis functions as values. This connection makes efficient attention implementations directly applicable to shape matching and an effective mechanism for scaling spatial-spectral coupling. We show that optimized quadratic attention implementations can scale modern deep functional map pipelines to large mesh sizes, improving both runtime and memory consumption, while sub-quadratic methods introduce errors, particularly for non-isometric shape pairs, where an exact coupling proves most necessary. These insights motivate Geo-FlashAttention, which restricts the exact softmax to geometry-selected surface patches and achieves near-linear empirical scaling while preserving correspondence accuracy of exact implementations. Across six standard benchmarks and three functional-map pipelines, rigorous analysis shows that efficient couplings preserve correspondence accuracy. At 60k vertices, these insights reduce peak training memory from 106.08 to 5.67,GiB, with Geo-FlashAttention being 1.5× faster per training step than FlashAttention. This enables end-to-end learning for correspondence problems on large meshes such as cortical surfaces, where high resolution is required to preserve disease-relevant intricacies of the surface geometry. Project page available https://spatial-spectral-attention.github.io/
Figures & tables
Figure 1: Scalability challenges of Deep Spectral Methods. Spectral matching methods often rely on recovering a point-to-point map ΠNM , either to convert a functional map back to the spatial domain or as an effective regularization. Comparing shape vertex features requires O(nM⋅nN) operations, which is infeasible with current strategies.
Figure 2: Spatial-spectral coupling common in functional map methods [ 12 ] . Shared encoders extract feature descriptors of both shapes, which are used to construct functional and point-to-point maps coupled in an alignment regularization.
Figure 3: Overview of Geo-FlashAttention. (1) Target N and source M are partitioned into geometrically derived patches A and B based on vertex positions. (2) Each source patch Br is summarised by its mean key μr . (3) Target queries are scored against the pooled keys and reduced per patch pair by sar=maxi∈Aaqi⊤μr . (4) The R best-scoring source patches per target patch form the support Sa,R , so every row keeps exactly R patches. (5) A fused block-sparse kernel evaluates the exact softmax without materialising ΠNM(R) .
Figure 4: Comparing attention computations. (1) Dense softmax forms the full nN×nM correspondence matrix, while FlashAttention (2) tiles A through on-chip memory. (3) Linear attention replaces the softmax kernel by a feature map φ approximating the product. (4) Geo-FlashAttention evaluates exact softmax attention on the induced vertex support by S .
Method
Coupling
FAUST
SCAPE
SHREC’19
SMAL
DT4D-intra
DT4D-inter
Iterative refinement
ZoomOut [ 46 , 44 ]
Dense
1.79±0.00
4.33±1.38
6.17±0.13
23.48±3.72
1.94±0.04
5.17±0.18
FlashAttention
1.79±0.00
4.18±1.53
5.70±0.67
25.51±2.81
1.87±0.15
5.13±0.15
Geo-FlashAttention
1.80±0.01
4.48±2.11
5.71±0.24
24.29±1.73
2.03±0.25
6.48±2.18
Learned pipelines
ULRSSM [ 12 ]
Dense
1.56±0.04
1.87±0.02
4.45±0.98
4.14±0.35
0.96±0.05
4.12±0.33
Table 1: Correspondence accuracy of feature couplings, as mean geodesic error ( ×100 , ↓ ) over three seeds ± standard deviation. Within each method block, only the computation of ΠΦ changes. Learned-pipeline rows differ due to coupling and optimization convergence. We run Bonferroni-corrected simultaneous 95% Welch confidence intervals for mean differences from dense runs, for which no statistically significant differences are detected (they all include zero). Our proposed couplings are in bold.
TOPKIDS ( ≈8 – 12 k vertices)
TOPKIDS-HD ( ≈60 k vertices)
Coupling
Time (ms) ↓
Peak mem. (GiB) ↓
Geo. err. ↓
Time (ms) ↓
Peak mem. (GiB) ↓
Geo. err. ↓
Exact couplings
Dense (fp32)
66.0±7.9
5.22
16.55±6.84 ( 5.92 )
710.6±105.6
106.08
OOM †
KeOps (fp32) [ 44 ]
1943.2±7.0
1.23
−
38285.9±115.5
5.67
Unsched. ∗
SDPA (fp32)
78.9±5.2
1.23
19.35±7.41 ( 8.43 )
861.5±5.0
5.67
11.05±1.69 ( 8.83 )
FlashAttention (bf16)
46.4±0.3
1.23
15.65±6.52 ( 9.08 )
207.2±5.0
5.67
11.38±3.11 ( 7.02 )
Table 2: Cost and accuracy of ULRSSM [ 12 ] couplings on the two released TOPKIDS datasets [ 35 ] . Time and peak memory are for one training step at ≈12 k and ≈60 k, mean ± standard deviation over five repeats; geodesic error ( ×100 ) uses 25 test pairs of the released datasets, mean ± standard deviation over fixed seeds with the best seed in parentheses (seed counts and protocol: Sec. 7.1 ). ∗ Unschedulable: 29.9 days for one 100 -epoch run. † Dense training at 60 k exceeds 140 GB; 106.08 GiB is one iteration.
Method
Geo. err. ↓
ZoomOut [ 46 ]
33.7
AttentiveFMaps [ 36 ]
23.4
Deep Shells [ 25 ]
13.7
Smooth Shells [ 24 ]
10.8
ULRSSM [ 12 ]
9.2
Synchronous Diffusion [ 13 ]
5.4
Table 3: Topological noise. Mean geodesic error ( ×100 ) on the released TOPKIDS benchmark [ 35 ] , over its 25 test pairs at their released resolution ( ≈8 – 12 k) and following existing protocols [ 12 , 6 ] . Published rows are quoted as reported. GeoFA uses Euclidean patches with an error of 4.33±0.42 over three seeds.
Label transfer
Darea↓ by regions S
Method
mIoU ↑
Acc. ↑
10
40
200
FreeSurfer ∗ [ 28 ]
0.718
0.851
0.0601
0.1129
0.1772
FBM [ 40 ]
0.619
0.795
0.0367
0.0783
0.2394
V2C-Flow [ 9 ]
0.568
0.753
0.0823
0.1192
0.2085
ULRSSM [ 12 ] + Geo-FlashAttention
0.633
0.800
0.0184
0.0325
0.0706
Table 4: Label transfer and area distortion on cortical surfaces , averaged over 90 validation pairs of the Mindboggle dataset. Left: manually corrected DKT31 parcellations as preserved by each map. Right: Darea(S) , a map distortion measure. S is the number of geodesic-Voronoi regions, and Darea growing with S . For Geo-FA, we report the mean over three seeds. All methods are run under identical settings. ∗ FreeSurfer [ 28 ] generated the reference labels prior to manual correction, so it is expected to score strongly on label transfer while confounding the evaluation w.r.t. parcelation labels.
Figure 5: Label transfer between two held-out Mindboggle-101 subjects (lateral pial view). Source DKT31 labels are transferred to the target. Disagreements with the target’s manually corrected labels are highlighted in red. Zoomed regions are depicted in numbered boxes. Percentages give vertex accuracy between the pairs. ∗ Reference labels were initialized with FreeSurfer before manual correction.
Figure S1 : Seed-dependent flip on SHREC’19 shape 28 . The source is coloured by signed lateral coordinate and pulled through the predicted map. One seed preserves the sides, while another exchanges them over the upper body.
Method
Coupling
SMAL
DT4D-inter
ZoomOut
Dense
23.23±4.02
8.03±2.49
FlashAttention
25.37±2.91
9.25±2.32
Geo-FlashAttention
24.20±1.95
11.22±0.85
ULRSSM
Dense
5.12±0.70
4.49±0.29
FlashAttention
4.64±0.69
6.17±2.26
Geo-FlashAttention
5.43±0.57
5.00±1.96
Table S1 : All-pairs results. Mean geodesic error ( ×100 , ↓ ) including the bistable pairs omitted from the corresponding SMAL and DT4D-inter cells of Tabs. 1 and S3 . The exclusion consistently improves the mean across learned pipelines. It does not improve ZoomOut on SMAL, whose errors are distributed broadly rather than concentrated on the bistable pairs.
ULRSSM / HFM
diff. ZoomOut
Benchmark
pairs
epochs
iters
TTA
pairs
iters
FAUST
6400
15
96000
0
3160
31600
SCAPE
2601
30
78030
0
1275
12750
SHREC’19
9001
10
90010
12
4435
44350
SMAL
841
20
16820
15
406
4060
DT4D-H
8879
10
88790
0 / 15
8702
87020
Table S2: Training schedule per benchmark, batch size one. SHREC’19 trains on FAUST and SCAPE; one DT4D training serves the intra- and inter-class evaluations ( 0 and 15 test-time adaptation (TTA) steps). Differentiable ZoomOut trains ten epochs everywhere, with no TTA beyond its ten refinement iterations, and is not run on TOPKIDS.
Figure S2 : Training time and peak memory across seven TOPKIDS resolutions, and accuracy relative to the dense coupling on the standard benchmarks. In (c), markers are means over five seeds, error bars their standard deviation, and the grey band the spread of the dense reference; the rows repeat the ULRSSM block of Tab. 1 and the ELU column of Tab. S3 under the same seeds and exclusions ( Sec. 7.1 ). Non-dense methods share a single memory curve. One representative per class of coupling is shown; Performer, KeOps and fp32 SDPA are omitted.
Dense
Linear attention
Benchmark
(reference)
ELU
Performer
FAUST
1.56±0.04
1.55±0.01
1.93±0.13∗∗∗
SCAPE
1.87±0.02
1.99±0.06∗∗
2.47±0.10∗∗∗
SHREC’19
4.45±0.98
4.97±0.46
5.64±1.42
SMAL
4.14±0.35
5.54±0.48∗∗∗
6.85±1.15∗∗∗
DT4D-intra
0.96±0.05
1.12±0.08∗∗
1.29±0.12∗∗∗
Table S3 : Cost of linearized couplings. Mean geodesic error ( ×100 , ↓ ); unlike Tab. 1 , these approximate the softmax kernel and change the correspondence. Markers: two-sample t -test against the dense reference over five seeds, ∗p<0.05 , ∗∗p<0.01 , ∗∗∗p<0.001 .
SVD
Nyström
top- k
spherical
Benchmark
support
m=200
m=1024
m=1024
k=128
p=32
FAUST
24.1
0.449
0.0209
0.149
0.0027
0.028
SCAPE
24.0
0.459
0.0133
0.160
0.0062
0.038
DT4D
36.0
0.587
0.0233
0.126
0.0113
0.041
SMAL
54.4
0.358
0.0058
0.157
0.0134
0.064
Table S4 : Approximating a converged coupling. Each method approximates the converged point-to-point coupling Π . We assess optimal rank- m truncation by SVD, Nyström with m landmarks, top- k row truncation, or the spherical kernel ( p=32 ). Entries report the relative Frobenius error of the functional map induced by the approximated coupling, for one converged checkpoint per benchmark. Support is the effective number of vertices per softmax row. Benchmarks are ordered by support; best per column in bold.
Figure S3 : Qualitative comparison of spatial–spectral couplings. Texture is transferred through maps estimated from the same shape pair and seed. The exact couplings preserve the same regions. On SMAL and DT4D-inter, Performer fragments the transferred grid, whereas ELU mainly shifts region boundaries, consistent with Tab. S3 . All couplings appear similar on FAUST.
Figure S4 : Learned cortical descriptors. One Mindboggle pair, colored by DKT regions, with the joint t-SNE embedding of its descriptors.
Figure S5 : Smooth-field transfer on Mindboggle-101. A source field is transferred to a held-out target using each correspondence method.
Figure S6 : Cortical parcellation transfer on Mindboggle-101. DKT31 source labels are transferred to the same target; the final panel shows its ground-truth atlas.
Figure S7 : Euclidean patches on self-contact regions in TOPKIDS. Euclidean patch partition of a TOPKIDS shape (kid06, 353 patches). (a, c) One patch highlighted in each; (b, d) close-ups of the boxed regions. Neither patch is a connected region: each covers two surface pieces that touch in R3 but are separated along the surface, the underside of the upper arm and the shoulder (b), and the two sides of an abdominal skin fold (d).
Table 21
Figure S8 : Geodesic error distributions. Cumulative fraction of correspondences below each error threshold, using the same runs and exclusions as Tabs. 1 and S3 . Near-isometric panels span 0 – 10 and the others 0 – 20 ; pair counts and exclusions are shown in each panel.
Deep functional maps, leveraging learned feature extractors and spectral correspondence solvers, are fundamental to non-rigid 3D shape matching. Based on an analysis of open-source implementations, we find that standard functional map implementations solve k independent linear systems serially, which is a computational bottleneck at higher spectral resolution. We thus propose a vectorized reformulation that solves all systems in a single kernel call, achieving up to a 33x speedup while preserving the exact solution. Furthermore, we identify and document a previously unnoticed implementation divergence in the spatial gradient features of the mainstay DiffusionNet: two variants that parameterize distinct families of tangent-plane transformations, and present experiments analyzing their respective behaviors across diverse benchmarks. We additionally revisit overlap prediction evaluation for partial-to-partial matching and show that balanced accuracy provides a useful complementary metric under varying overlap ratios. To share these advancements with the wider community, we present an open-source codebase, DeepShapeMatchingKit, that incorporates these improvements and standardizes training, evaluation, and data pipelines for common deep shape matching methods. The codebase is available at: https://github.com/xieyizheng/DeepShapeMatchingKit
Yizheng Xie, Lennart Bastian, Congyue Deng +3
Technical University of Munich · Simon Fraser University · Munich Center for Machine Learning +3
Finding dense correspondences between 3D shapes is a fundamental yet unresolved challenge, especially in real-world environments. These environments present severe challenges, including the lack of time and sufficient samples for training, the prevalence of uncurated extreme-high resolution data with topological distortions, and the need to handle diverse 3D representations. In this paper, we present ATM, a zero-shot framework that requires no correspondence-specific training and robustly addresses these issues at once through an articulate-then-match paradigm. Rather than relying on intrinsic geometric properties, we leverage powerful pretrained vision foundation models and parametric shape priors to estimate parametric shape models from multi-view renderings, and systematically ground these estimations via multi-view geometric consistency. By mapping diverse inputs into a shared canonical parametric space, we inherently establish robust coarse correspondences that bypass topological noise, which are then refined into precise dense mappings via spectral refinement. Operating purely on test-time optimized parametric reconstructions, ATM requires no correspondence training data, is naturally immune to connectivity artifacts, and seamlessly handles diverse 3D modalities, including meshes, point clouds, and 3D Gaussians. Extensive experiments demonstrate that our method achieves strong results on non-isometric benchmarks (average geodesic errors of 2.4-TOPKIDS, 3.8-SMAL), reducing errors by 73% and 37% respectively compared to the baseline URSSM. Furthermore, it exhibits unprecedented robustness on in-the-wild raw scans of up to 200k vertices per shape while maintaining near-constant computation time and consistent superior accuracy.
Qilong Liu, Qinfeng Xiao, Chenyuan Yi +2
Hong Kong Polytechnic University, HK SAR · Artificial Intelligence in Design, HK SAR
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.
Adeela Islam, Zorah Lähner, Vittorio Murino +1
1Italian Institute of Technology · University of Genoa · 6Max Planck Institute for Informatics, SIC +4