---
title: 'TexParts: 3D Part Segmentation Benchmark'
url: https://www.emergentmind.com/topics/texparts
type: topic
---

# TexParts: 3D Part Segmentation Benchmark

Tex-Parts refers to a large-scale, affordance-grounded 3D part segmentation and naming benchmark constructed to address gaps in semantic 3D part annotation, vocabulary consistency, and evaluation. It is introduced in conjunction with the ALIGN-Parts method, which directly aligns geometric decompositions of 3D shapes with a unified ontology of canonical part labels, providing human-verified, semantically meaningful annotations at scale [2512.18003].

## 1. Scope, Ontology, and Dataset Construction

Tex-Parts is designed around a unified 3D part ontology comprising 1,794 canonical categories. The vocabulary was consolidated by aligning and compressing the original labels from three major 3D shape benchmarks—PartNet (32,141 shapes, ~2,348 original labels), 3DCoMPaT++ (8,627 shapes, ~1,468 labels), and Find3D (124 shapes, 640 labels). This compression process involves MPNet embedding similarity calculations to generate label merger candidates, Gemini 2.5 LLM verification to validate candidate merges, and human curation with preference for descriptiveness and preservation of distinct semantics (e.g., “car_front_bumper” vs. “car_rear_bumper” are maintained as separate classes). An alias mapping is preserved for backward compatibility and ontology expansion operations.

The resulting Tex-Parts dataset, bootstrapped from TexVerse and annotated via ALIGN-Parts, contains 8,450 rigorously verified 3D meshes. These include approximately 100,000 total part instances, with each object averaging 12 labeled parts (standard deviation ~4). The category distribution is long-tailed, with top-10% most frequent parts representing over 45% of all occurrences, and roughly 30% of the canonical labels appearing in five or fewer shapes. Tex-Parts is split into a large training set covering all three sources (40,982 shapes) and a held-out test set (206 shapes, with contributions from each benchmark and some novel Find3D examples).

## 2. Annotation Pipeline and Part Semantics

Tex-Parts annotation follows a confidence-driven human-in-the-loop pipeline powered by ALIGN-Parts. Candidate shapes from TexVerse are first pre-filtered by Gemini LLM to exclude malformed meshes. ALIGN-Parts then produces, in a single inference pass, a set of part masks, predicted labels, and per-part confidences. For high-confidence predictions (confidence ≥ 0.8), segmentations are auto-accepted. Shapes or parts with low confidence scores (< 0.5), or ambiguous output, are sorted for priority human review.

Annotation proceeds in two phases: (1) minor edits and corrections (bounding, mask refinement, or relabeling); and (2) annotation of missing or previously unmodeled part categories, discovered via canonical list search or LLM suggestion. To increase semantic clarity and disambiguate similar terms, each part is paired with an affordance description (e.g., “chair_seat” → “the horizontal surface where a person sits”), produced by Gemini 2.5 LLM and embedded with MPNet. This long-form descriptor is employed during both training and open-vocabulary inference, enabling disambiguation between functionally similar but contextually distinct part types (such as mug handle vs. door handle).

## 3. Methodology and Model Architecture

The ALIGN-Parts architecture decomposes each 3D shape into K=32 “partlets”—learnable soft masks associated with semantic embeddings in MPNet text space and a partness score. The overall process fuses geometric features (from a frozen PartField backbone) and appearance features (via multi-view DINOv2 encodings) by local, cross-modal attention (BiCo fusion) with 3D Fourier positional bias. The resulting fused point features inform the partlet decoding transformer stack, which refines semantic embeddings, predicts soft masks, and determines per-partlet activation.

Partlet-to-annotation matching is performed by optimal transport. The cost function balances geometric overlap (one minus Dice coefficient between partlet and ground truth mask) and semantic embedding similarity (one minus cosine similarity between partlet and ground-truth part description embedding). Matching is solved with Sinkhorn-regularized soft assignment followed by hard assignment. Supervision includes label alignment (InfoNCE loss on partlet–text pairs), mask and partness losses (binary cross entropy and Dice), regularization for part coverage and overlap, and a shape–class contrastive loss for closed-vocab shape recognition. The full training objective is a weighted sum:
$$
L_{total} = \lambda_{mask} L_{mask} + \lambda_{part} L_{part} + \lambda_{text} L_{text} + \lambda_{cov} L_{cov} + \lambda_{ov} L_{overlap} + \lambda_{global} L_{global}
$$

## 4. Evaluation Metrics

Tex-Parts introduces three principal metrics for named 3D part segmentation:

| Metric                  | Definition                                                                                              | Label Awareness              |
|-------------------------|---------------------------------------------------------------------------------------------------------|------------------------------|
| Class-agnostic mIoU     | Mean IoU between ground-truth part masks and best-overlapping predicted mask (ignores labels)           | None                         |
| Label-Aware mIoU        | IoU for matches where predicted label exactly matches ground truth; 0 otherwise                          | Strict (exact)               |
| Relaxed LA-mIoU         | Weights IoU by MPNet cosine similarity between predicted and GT part descriptions                        | Partial (semantic similarity)|

By construction: class-agnostic mIoU ≥ relaxed LA-mIoU ≥ LA-mIoU. These metrics explicitly quantify both geometric and semantic correspondence, supporting both closed- and open-vocabulary segmentation evaluation.

## 5. Quantitative Results and Comparative Analysis

On the unified held-out test set comprising 3DCoMPaT++, PartNet, and Find3D samples, ALIGN-Parts achieves a class-agnostic mIoU of 0.600, strict label-aware mIoU of 0.316, and relaxed label-aware mIoU of 0.529. These figures represent a 59% gain in LA-mIoU over the best baseline and a 100-fold inference speedup compared to PartField+K-means. Heavy-tailed category distributions are effectively managed by the ontology alignment, and annotation efficiency is increased by confidence calibration: Mahalanobis-Softmax–fused confidences enable auto-accepting high-confidence predictions, reducing manual annotation time by 5–8×.

The protocol further enables zero-shot (open-vocabulary) segmentation by bipartite assignment between predicted partlets and arbitrary LLM-derived part descriptions, subject to the semantic range modeled by the training data.

## 6. Downstream Applications and Limitations

Tex-Parts, in concert with ALIGN-Parts, functions as a scalable annotation engine for new 3D assets and enables ontology construction workflows that are extensible to novel domains. It supports text-conditioned part retrieval and label transfer for co-segmentation tasks. The affordance-based descriptions and canonical vocabulary facilitate functionally-grounded, semantically precise matching across shape collections.

Notable limitations include coverage restricted to the 1,794 trained part categories, reliance on frozen visual and geometric backbones (limiting adaptability across domains), and partial degradation of confidence calibration under significant domain shift or on noisy 3D scans. Open-vocabulary transfer remains constrained to “functionally similar” categories encountered during training. Directions for future work include joint learning of 3D backbones, self- and weak-supervision regimes, and integration with generative 3D priors and articulation modeling.

## 7. Relationship to Related Benchmarks

Tex-Parts establishes a new standard for semantically rich, large-scale 3D part segmentation with rigorous label grounding and cross-dataset vocabulary alignment. It advances beyond prior datasets—often limited by inconsistent label definitions or lack of semantic alignment—by providing affordance-grounded, human-verified part names across a unified ontology. The benchmark addresses both geometric accuracy and semantic consistency, supplying metrics, data, and model architectures to enable further research in 3D understanding, annotation automation, and shape-conditioned generative modeling [2512.18003].

Source: https://www.emergentmind.com/topics/texparts