SFD-Net: Sharp Feature Detection Network Based on Local Geometric Features

Detection fails where a representation gets the geometry wrong. Find that relation, then supply the explicit cue that fixes it.

Here the relation is the discontinuity of the surface-normal field, and the cue is a small multi-scale descriptor. The descriptor, not any backbone-specific tuning, is what drives the gains, which is why it carries across architectures and from synthetic CAD to real scans.

Abstract

Sharp features are the loci where a surface's normal field is discontinuous: the creases and corners that define an object's structure. They are also where detection breaks down, because joint density variation and noise destabilize the local statistics detection relies on, and because sharp points are a small minority class. SFD-Net addresses this with the Local Geometric Descriptor (LGD): per-point normals estimated at three nested neighborhood scales, aggregated as second-moment statistics of normal differences into a compact, scale-aware isotropy cue. Because it operates on normal differences, LGD is invariant to rigid motion, global scale, and normal sign flips, with no explicit alignment. It prepends to any point-cloud backbone without modification: on the ABC benchmark it reaches state-of-the-art F1 across four density-and-noise conditions, improves three independent backbones in a plug-and-play setting, and transfers zero-shot from synthetic CAD to real scans up to three orders of magnitude larger.

Transferable

A descriptor, not an architecture

LGD prepends to PointNet++, RepSurf, and EDWG unmodified, adding +4.05 to +7.00 F1 while cutting FPR by 2.62–9.08. The descriptor drives the gains.

State of the art

Leads ABC across every condition

Best F1 in every one of the four density-and-noise settings (61.33 → 57.05), including the largest-noise setting.

By construction

Invariant without alignment

Operating on normal differences makes LGD invariant to rigid motion, global scale, and normal sign flips, with no orientation propagation.

Generalizes

Zero-shot, CAD → real

Trained only on 10K-point CAD, applied directly to real S3DIS scans up to 7.1M points (~710×) with no retraining or subsampling.

Method

The Local Geometric Descriptor

LGD turns each point into a three-value cue, then hands it to a learned backbone. The geometry is made explicit before learning, rather than left for the network to infer from raw coordinates.

SFD-Net overview: LGD generation followed by an enhanced PointNet++ — **Two stages.** LGD is computed from multi-scale normal statistics, concatenated with raw coordinates, and fed to an enhanced PointNet++ with two lightweight Transformer blocks and an imbalance-aware loss.

Multi-scale PCA normals

Estimate a unit normal at three nested neighborhoods (k = 20, 40, 80). Only directions are kept, giving sign-flip invariance.

Second moments of normal differences

Accumulate weighted outer products of normal differences per scale. Using Δn Δn⊤ yields rigid-motion and global-scale invariance with no alignment.

Per-scale isotropy index

Collapse each scale to one isotropy value, low on planar patches and high near creases and corners, and concatenate the three into a compact cue.

What LGD is, precisely. It is a precomputed geometric cue that complements a learned backbone, not a replacement for learned features, and the pipeline is not end-to-end differentiable. Because the gains come from the descriptor, it transfers across backbones rather than being tied to one. Replacing the PCA step with a learned estimator is left for future work.

Results

State of the art on ABC

Four conditions: density variation only (none), then increasing Gaussian noise. SFD-Net leads F1 in every condition while keeping FPR among the lowest.

Average F1 ↑ / FPR ↓ on ABC (%). Bold row: ours.
Method	none F1	FPR	0.12% F1	FPR	0.6% F1	FPR	1.2% F1	FPR
PointNet++	50.89	48.57	55.02	45.88	49.03	49.35	50.21	49.01
DGCNN	49.89	48.90	50.65	48.51	48.44	49.58	47.53	49.99
RepSurf	55.23	45.96	57.16	43.06	54.21	46.55	48.17	49.73
PointMLP	54.59	46.30	47.53	50.00	47.52	50.01	47.65	49.94
PIE-Net	52.55	38.50	52.44	37.35	47.61	40.54	49.56	39.33
BoundED	48.12	49.88	47.98	49.94	47.97	49.90	47.70	49.98
SFC-Net	47.92	49.84	48.64	49.53	48.27	49.69	48.03	49.80
MSL-Net	59.41	–	59.11	–	56.11	–	52.74	–
EdgeFormer	52.04	31.24	48.66	33.95	34.99	45.16	28.13	48.55
EDWG	54.95	46.06	54.93	46.07	48.20	49.81	47.76	49.90
SFD-Net (ours)	61.33	36.93	60.42	39.76	58.62	40.88	57.05	41.94

Qualitative comparison on ABC — **ABC (synthetic CAD).** EdgeFormer over-detects on smooth faces and EDWG misses structure, while SFD-Net stays close to ground truth.

Architecture-agnostic

Prepending LGD, with no architectural change, improves all three backbones. LGD + EDWG (60.80 F1) nearly matches full SFD-Net (61.33), confirming the descriptor is the primary driver.

Plug-and-play LGD on ABC (density variation only), %.
Backbone	F1 ↑	FPR ↓	ΔF1	ΔFPR
PointNet++	50.89	48.57	—	—
+ LGD	57.89	43.70	+7.00	−4.87
RepSurf	55.23	45.96	—	—
+ LGD	59.28	43.34	+4.05	−2.62
EDWG	54.95	46.06	—	—
+ LGD	60.80	36.98	+5.85	−9.08

Zero-shot transfer to real scans

Trained only on 10K-point synthetic CAD, SFD-Net is applied directly to real S3DIS scenes (363K to 7.1M points) with no retraining, fine-tuning, or subsampling.

Zero-shot sharp feature detection on a real S3DIS room — **Real S3DIS room.** EDWG floods flat surfaces with false positives; SFD-Net recovers clean wall, partition, and furniture edges.

Citation

BibTeX

Placeholder — to be replaced with the official ECCV 2026 entry once released.

@inproceedings{oh2026sfdnet,
  title     = {SFD-Net: Sharp Feature Detection Network Based on Local Geometric Features},
  author    = {Oh, Inyoung and Ko, Kwang Hee},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}