Unified Primitive Proxies for Structured Shape Completion

TL;DR

UniCo introduces a unified, structured approach to 3D shape completion that jointly predicts assembly-ready quadratic primitives and completed points from partial scans. By using primitive proxies—learnable queries that attend to shared shape features—and coordinating two pathways, UniCo produces primitives with complete geometry, semantics, and inlier membership in a single pass, enabling robust assembly-based reconstruction. The training employs online target updates and permutation-invariant matching to maintain stable, self-consistent supervision as predictions evolve. Across synthetic and real datasets, UniCo achieves up to 50% lower Chamfer distance and up to 7% higher normal consistency than recent baselines, demonstrating the practical benefits of coordinated completion and primitive inference for structured 3D understanding.

Abstract

Structured shape completion recovers missing geometry as primitives rather than as unstructured points, which enables primitive-based surface reconstruction. Instead of following the prevailing cascade, we rethink how primitives and points should interact, and find it more effective to decode primitives in a dedicated pathway that attends to shared shape features. Following this principle, we present UniCo, which in a single feed-forward pass predicts a set of primitives with complete geometry, semantics, and inlier membership. To drive this unified representation, we introduce primitive proxies, learnable queries that are contextualized to produce assembly-ready outputs. To ensure consistent optimization, our training strategy couples primitives and points with online target updates. Across synthetic and real-world benchmarks with four independent assembly solvers, UniCo consistently outperforms recent baselines, lowering Chamfer distance by up to 50% and improving normal consistency by up to 7%. These results establish an attractive recipe for structured 3D understanding from incomplete data. Project page: https://unico-completion.github.io.

Figures (8)

• Figure 1: We present UniCo, a structured shape completion model that, given a partial scan, jointly predicts a complete set of quadratic primitives with geometry, semantics, and inlier membership. The predicted primitives are assembly-ready for surface reconstruction.
• Figure 2: Architecture of UniCo. Shape features from a partial point cloud feed two coordinated pathways. The point pathway decodes dense completed points. The primitive pathway uses primitive proxies that attend to the shared features and predict primitive semantics, geometry, inlier membership, and a confidence score used at inference to select valid primitives. The selected primitives are assembly-ready.
• Figure 3: Comparison to completion baselines on ABC-multi. For each baseline, completed points are paired with its best-performing primitive extractor ( RANSAC schnabel2007efficient, HPNet yan2021hpnet, PTv3 wu2024ptv3). UniCo recovers extractor-free, assembly-ready primitive structures.
• Figure 4: Comparison with reconstruction methods. Even with completed points with better pointwise metric, both competitors cannot produce detailed and robust reconstructions.
• Figure 5: Robustness to missing data, transforms, and noise. Left: as incompleteness grows from 25% to 75%, UniCo maintains lower CD and higher NC than pointwise baselines. Middle: under the debiased normalization protocol wang2025simeco, UniCo remains stable with low pose and scale bias compared to baselines. Right: under Gaussian jitter of 1--3%, performance degrades gracefully.