PatchAlign3D: Local Feature Alignment for Dense 3D Shape understanding

2026-01-05

arXiv.org

AI Summary

This work addresses the limited transferability of existing 3D foundation models in local part understanding tasks, which often rely on multi-view rendering and large language model (LLM) prompting while neglecting intrinsic 3D geometric structure. The authors propose a novel encoder-only 3D point cloud model that, through a two-stage pretraining strategy, directly generates local features aligned with textual semantics—enabling, for the first time, zero-shot part segmentation via a single forward pass without multi-view rendering or LLM prompts. Built upon a point cloud Transformer, the method integrates DINOv2-based dense feature distillation with multi-positive contrastive learning to effectively align 3D local geometry with part-level text embeddings. Experiments demonstrate that the model significantly outperforms current approaches across multiple 3D part segmentation benchmarks, achieving both high efficiency and accuracy.

Technology Category

• Computer Vision: 3D Computer Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Current foundation models for 3D shapes excel at global tasks (retrieval, classification) but transfer poorly to local part-level reasoning. Recent approaches leverage vision and language foundation models to directly solve dense tasks through multi-view renderings and text queries. While promising, these pipelines require expensive inference over multiple renderings, depend heavily on large language-model (LLM) prompt engineering for captions, and fail to exploit the inherent 3D geometry of shapes. We address this gap by introducing an encoder-only 3D model that produces language-aligned patch-level features directly from point clouds. Our pre-training approach builds on existing data engines that generate part-annotated 3D shapes by pairing multi-view SAM regions with VLM captioning. Using this data, we train a point cloud transformer encoder in two stages: (1) distillation of dense 2D features from visual encoders such as DINOv2 into 3D patches, and (2) alignment of these patch embeddings with part-level text embeddings through a multi-positive contrastive objective. Our 3D encoder achieves zero-shot 3D part segmentation with fast single-pass inference without any test-time multi-view rendering, while significantly outperforming previous rendering-based and feed-forward approaches across several 3D part segmentation benchmarks. Project website: https://souhail-hadgi.github.io/patchalign3dsite/

Problem

• 3D part segmentation
• local feature alignment
• dense 3D understanding
• foundation models
• point cloud

Innovation

• PatchAlign3D
• 3D part segmentation
• point cloud transformer
• feature distillation
• contrastive alignment