P3-SAM: Native 3D Part Segmentation

• P3-SAM produces precise part segmentation results for any object.Source: Figure 1 · p. 1
• Figure 1 is associated with the paper’s reported segmentation result.Source: Figure 1 · p. 1
• The supplied author note identifies a project leader and corresponding author.Source: §1 · p. 1

• Existing methods struggle with arbitrary parts because category-specific supervision limits generalization, while 2D-lifting methods suffer data gaps, imprecision, and incomplete automation.Some methods additionally require the number of parts or user-provided prompt points.Source: §1 · p. 2
• P3-SAM uses a native 3D point-promptable architecture with a feature extractor, three segmentation heads, and an IoU prediction head.A single positive point prompt indicates a part, and three multi-scale masks are evaluated to select the best one.Source: §1 · p. 2
• The method automatically samples prompts, filters redundant masks, and projects point-level masks onto mesh faces for part segmentation.Source: §1 · p. 2
• Training on 3.7 million meshes with high-quality part masks enables robust, precise, and globally coherent segmentation without relying on 2D SAM.Source: §1 · p. 2
• P3-SAM achieves state-of-the-art part segmentation for arbitrary objects, especially complex objects with highly detailed geometry.The result is shown in Figure 1.Source: §1 · p. 2
• The method is reported to support interactive, multi-head, and hierarchical part segmentation across varied tasks and data types.Source: §1 · p. 2

2.1 TRADITIONAL 3D PART SEGMENTATION

• Traditional 3D methods use predefined labels and 3D encoders but struggle to generalize beyond specific categories and parts.Source: §2.1 · pp. 2–3

2.2 2D LIFTING 3D PART SEGMENTATION

• 2D lifting methods aggregate multi-view SAM or VLM outputs for class-agnostic 3D segmentation, but face data gaps, 3D inconsistency, unstable post-processing, and high resource costs.Source: §2.2 · p. 3

2.3 2D DATA ENGINE FOR 3D PART SEGMENTATION

• 2D data engines distill or supervise 3D networks using projected foundation-model features or masks, improving generalization while retaining boundary ambiguities and data gaps.Source: §2.3 · p. 3
• These 2D-data approaches still require specified category counts or user prompt points, preventing fully automatic object segmentation.Source: §2.3 · p. 3
• Figure 2 depicts P3-SAM’s feature extractor, two-stage multi-mask segmentor, and IoU predictor for selecting among three mask scales.Source: Figure 2 · p. 4

3 METHOD

• P3-SAM predicts an instance-specific, class-agnostic part label for every mesh face, with the number of parts determined per object.The target mask assigns each face to one of Npart parts.Source: §3 · p. 4

3.1 DATA CURATION

• Nearly 3.7 million filtered objects and nearly 2.3 million watertight models provide training data for both non-watertight and watertight geometry.Watertight versions are sampled with 80% probability when available.Source: §3.1 · p. 4

3.2.1 NETWORK ARCHITECTURE

• A pretrained Sonata encoder extracts reusable multi-scale point features, while the point features are computed once for segmentation with different prompts.The extracted features combine point coordinates, normals, and multi-scale encoder outputs.Source: §3.2.1 · pp. 4–5
• Three first-stage heads predict alternative masks from point features and a single point prompt, then a second stage uses global features and initial masks to refine them.The multi-head design addresses varying part granularity and ambiguous prompt scale; initial masks also focus global-feature extraction on relevant parts.Source: §3.2.1 · p. 5
• An IoU predictor selects the best of three refined masks, enabling lightweight real-time interactive segmentation after global features are extracted.It predicts three IoU values from global features and the second-stage masks.Source: §3.2.1 · p. 5

3.2.2 TRAINING

• Training adds noise to points, normals, and prompts, randomly removes normals with probability 0.3, and mixes watertight with non-watertight data.These augmentations are introduced to enhance robustness.Source: §3.2.2 · p. 5
• For each sampled part prompt, Dice and focal losses supervise both stages, while backpropagating only through the lowest-loss head encourages different heads to learn different scales.The total objective combines both stage mask losses with IoU supervision based on mean squared error.Source: §3.2.2 · pp. 5–6

3.3 AUTOMATIC SEGMENTATION

• Automatic segmentation oversamples FPS point prompts, predicts masks and IoUs, suppresses redundant masks with NMS, and projects selected point masks onto mesh faces.Face labels are assigned by voting, and flood fill labels remaining unlabeled faces from neighboring faces.Source: §3.3 · p. 6, Figure 3 · p. 6, Table 3 · p. 7

4.1 COMPARISON

• Interactive evaluation samples 10 prompt points per part and averages predicted-mask IoU against corresponding ground-truth masks.Source: §4.1 · p. 8
• Quantitative and qualitative comparisons across non-watertight meshes, watertight meshes, and point clouds demonstrate effectiveness, robustness, and generalization across three segmentation tasks.The evaluation compares P3-SAM with recent methods on PartObj-Tiny, PartObj-Tiny-WT, and PartNetE.Source: §4.1 · pp. 7–8

4.2 APPLICATIONS

• Multi-prompt segmentation follows user instructions by extending unsegmented regions and merging over-segmented regions.Source: §4.2 · p. 8, Figure 5 · p. 8
• Hierarchical segmentation aggregates parts at multiple levels and adheres better to part relationships than PartField.Source: §4.2 · pp. 8–9
• More accurate P3-SAM masks improve HoloPart exploded-part generation, producing cleaner and more precise parts than masks from SAMPart3D.Source: §4.2 · p. 9

4.3 ABLATION STUDY

• Ablations show that the IoU predictor, two-stage multi-head segmentor, global feature, data augmentation, NMS, and flood fill each contribute to complete and reasonable segmentation.Without NMS or flood fill, masks have incomplete coverage, unclear boundaries, and an unreasonable number of parts; global features improve the third ablated model over the second.Source: §4.3 · p. 9, Table 4 · p. 9, Figure 6 · p. 9

• P3-SAM uses Sonata point-wise features, a two-stage multi-head segmentor, and an IoU predictor to select multi-scale masks from point prompts.The method is trained on 3.7 million models and includes an automatic segmentation approach.Source: §5 · p. 9
• The method provides high accuracy, generalization, and robustness across various tasks and data types, including interactive, hierarchical, and multi-prompt segmentation.Source: §5 · pp. 9–10
• Because training uses only surface point clouds, P3-SAM may rely too heavily on surface geometry and lack understanding of parts’ spatial volume.Future work is proposed to develop spatial segmentation capabilities.Source: §5 · p. 10

• The paper attributes all technical contributions, including methodology, equations, and results, solely to the authors.Source: §A.1 · p. 12

• The section provides a more detailed introduction to related works and highlights differences between existing methods and the proposed method.Source: §A.2 · p. 12

• Traditional methods use 3D encoders and segmentation heads trained with predefined part labels for specific object or scene categories.Source: §A.2.1 · p. 12
• Their dependence on detailed category labels makes large-scale annotation impractical and weakens generalization to arbitrary categories and parts.Source: §A.2.1 · p. 12

• 2D lifting methods render 3D assets into multiple views, apply models such as SAM or GLIP, and aggregate or post-process the results for 3D part segmentation.Source: §A.2.2 · p. 13
• These approaches can suffer from 2D–3D data gaps, inconsistent geometry, unstable post-processing, prompt-engineering requirements, and substantial rendering cost.Source: §A.2.2 · p. 13

A.2.3 2D DATA ENGINE FOR 3D PART SEGMENTATION

• P3-SAM is trained exclusively on native 3D data, avoiding 2D data engines while using substantially more data to learn geometric features.Source: §A.2.3 · pp. 13–14
• Unlike Point-SAM, P3-SAM uses a 3D-specific feature extractor, a single point prompt, and an automatic segmentation method rather than manual prompt selection.Source: §A.2.3 · p. 14

A.3.1 DATA CURATION DETAILS

• Nearly 3.7 million objects are curated by decomposing meshes, merging parts below 1% area, filtering objects with 2–50 parts, and removing severe area imbalance; nearly 2.3 million watertight versions are also produced.Source: §A.3.1 · pp. 14–15

A.3.2 DATA AUGMENTATION DETAILS

• Noise augmentation randomly samples scales up to 0.01 and perturbs points, normals, and prompt locations to simulate varying point-cloud noise.Source: §A.3.2 · pp. 15–16

A.3.3 AUTOMATIC SEGMENTATION DETAILS

• Automatic segmentation samples points and prompts, predicts masks and IoU values, removes masks with IoU overlap above 0.9, assigns retained masks to the mesh, and flood-fills unlabeled faces.Source: §A.3.3 · p. 16

A.4 EVALUATION DETAILS

• Across non-watertight meshes, watertight meshes, and point clouds, P3-SAM consistently achieves the best reported performance and maintains segmentation quality when connectivity information is absent.Source: §A.4 · p. 17
• Qualitative results show accurate boundaries, reasonable part counts, and robust handling of complex objects and varied data types, unlike baselines that over-segment or produce inaccurate masks.Source: §A.4 · p. 17

A.5 APPLICATIONS DETAILS

• The model supports multi-prompt auto-segmentation, hierarchical clustering of part features, interactive real-time segmentation, and improved part-generation inputs.Source: §A.5 · pp. 17–18