Research map / Point clouds / Point cloud classification

Riemann-based Multi-scale Attention Reasoning Network for Text-3D Retrieval

AAAI 2025 · Point clouds › Point cloud classification · approach: Point-language and open-vocabulary 3D · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RMARN, a Riemann-geometry-based multi-scale attention network for text-to-3D point cloud retrieval, and releases the T3DR-HIT dataset of about 3,380 text-3D pairs.

Problem

Text-3D point cloud retrieval is underexplored due to scarce paired text-3D data and the irregular structure of point clouds, and Euclidean cosine similarity fails to capture their geometric relations.

Approach

Treats text and point cloud token sequences as fields on learned Riemannian manifolds, computing token-level similarity via a Riemann Attention Mechanism (RLS), fused with a global pooled cosine similarity (GPS) and a low-rank filter, without explicitly defining the manifold.

Key results

• Best configuration (BERT + PointNet++, 8 SA layers) reaches R@1 31, R@5 61, R@10 69, Rsum 161 on T3DR-HIT.
• Removing GPS drops Rsum from 161 to 75 in the ablation study.
• Removing both AFR and RLS yields Rsum 140; removing only RLS yields Rsum 148.

Measured effects reported in the paper

Surprising findings

• Removing the GPS module causes the largest performance drop (Rsum 161 to 75), larger than removing RLS or AFR.

Limitations

• Evaluation only on the authors' own custom datasets; no comparison with established text-3D retrieval baselines.
• Riemannian manifold formulation is approximated and loosely justified; the manifold parameters are learned rather than defined.
• Ablation and hyperparameter results are inconsistent in the paper (e.g. Table 1 labels and text about Nhead and SA layers differ).

Experimental setup

Datasets: Stanford 2D-3D-Semantics Elephant Meta Dataset

Benchmarks: T3DR-HIT

Metrics: R@1 R@5 R@10 Rsum

Baselines: GNN-based PointNet GNN-based PointNet++ CLIP-based RMARN variants BERT-based RMARN variants

Builds on: CLIP PointNet contrastive loss (CLIP-style)

Closest papers by meaning

• More Text, Less Point: Towards 3D Data-Efficient Point-Language Understanding (AAAI 2025 · similarity 0.86)
• Learning from Videos for 3D World: Enhancing MLLMs with 3D Vision Geometry Priors (NeurIPS 2025 · similarity 0.85)
• MLLM-For3D: Adapting Multimodal Large Language Model for 3D Reasoning Segmentation (NeurIPS 2025 · similarity 0.84)
• Merge3D: Efficient 3D Multimodal LLMs via Joint 2D-3D Token Merging (CVPR 2026 · similarity 0.84)
• Retrieval-Augmented Score Distillation for Text-to-3D Generation (ICML 2024 · similarity 0.84)
• Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning (CVPR 2026 · similarity 0.84)