Talk2Point

Granular Region-Specific Text-Guided 3D Point Cloud Editing

Organization: Independent Research

Key Result & Impact Benchmark

Achieved superior region-specific 3D shape editing accuracy compared to baseline global shape models.

Technical Overview

Talk2Point is a side research project exploring fine-grained 3D shape modification guided by natural language comparisons. By extending the ChangeIt3D architecture on the ShapeTalk dataset, Talk2Point replaces global shape tokens with localized 3D point cloud patch embeddings. This allows the network to interpret fine-grained spatial instructions (such as 'make the chair armrests thicker' or 'curve the backrest') and modify only the relevant 3D regions while preserving the rest of the shape geometry.

Key Technical Highlights

Modified ChangeIt3D tokenization to extract multi-scale 3D patch embeddings for localized spatial alignment.
Evaluated text-driven shape editing performance on ShapeTalk benchmark pairs.
Integrated SDFusion backends for zero-shot 3D point cloud reconstruction from text prompts.

Technologies & Frameworks

ShapeTalk Dataset

ChangeIt3D

SDFusion

Open3D

PyTorch