---
annotations_creators:
- expert-generated
language:
- en
- zh
tags:
- 3d-segmentation
- remote-sensing
- lidar
- point-cloud
- dino
- urban-modeling
datasets:
- SensatUrban
- STPLS3D
---

# DPT (Decoupled Point Transformer) Dataset & Weights

This repository contains the processed datasets and pre-trained checkpoints for **DPT (Decoupled Point Transformer)**, an advanced 3D semantic segmentation framework built on PTV3, integrating deep 2D visual priors from DINO.

Developed by **Bole Zhang (University of Bristol)**.

## 🚀 Overview

DPT addresses the semantic gap in 3D Point Clouds by decoupling geometric features and deep 2D cognitive priors. This dataset includes our **1031D (or 1025D) aligned features** specifically designed for Urban Remote Sensing.

### Key Features:
- **PTV3 Backbone**: Leveraging state-of-the-art Serialized Attention.
- **DINOv2/v3 Integration**: Each point is augmented with a 1024-dimensional visual descriptor.
- **GCDM Module**: Geometric-Cognitive Decoupling Module for dynamic modal fusion.
- **Top Performance**: Achieved SOTA on SensatUrban and STPLS3D datasets.

## 📦 Data Structure

The processed data is stored in `.npy` or `.pth` chunks (50m stride for SensatUrban). Each data point contains:

| Feature Name | Dims | Description |
| :--- | :--- | :--- |
| `coord` | 3 | Normalized XYZ coordinates |
| `color` | 3 | RGB values (0-255) |
| `rel_z` | 1 | Global Relative Elevation (Physical Prior) |
| `dino_feat` | 1024 | Deep semantic priors from DINOv2/v3 |
| **Total** | **1031** | **Pure Aligned Feature Vector** |

## 🛠️ Usage

### Environment Setup
This project requires the **Pointcept** framework and `spconv 2.x`.
```bash
# Clone the repository
git clone [https://github.com/zbole/DPT.git](https://github.com/zbole/DPT.git)
cd DPT