EyeNet++: A Multi-Scale and Multi-Density Approach for Outdoor Point Cloud Semantic Segmentation

Overview

EyeNet++ is a novel semantic segmentation network designed for outdoor 3D point cloud datasets. Inspired by the human vision field, EyeNet++ introduces a multi-scale and multi-density input framework combined with parallel processing architecture. These innovations enable enhanced coverage and accurate feature extraction for dense point cloud data.

Key Features:

• Multi-Scale and Multi-Density Input Scheme: Captures diverse spatial features using dense central and sparse peripheral regions.
• Dense Local Feature Aggregation (DLFA): Efficiently extracts contextual information from objects of varying sizes.
• Connection and Feature Merging Blocks: Facilitate effective exchange and integration of features across multi-scale inputs.
• Proven Performance: Demonstrated state-of-the-art results on benchmark datasets such as SensatUrban, Toronto3D, and YUTO Semantic.

Status

The source code for EyeNet++ and pre-trained models will be released soon. Stay tuned for updates!

Datasets

EyeNet++ has been extensively evaluated on several large-scale datasets:

• SensatUrban
• Toronto3D
• YUTO Semantic
• DALES
• Semantic3D

Upcoming Release

The GitHub repository will include:

1. Implementation code for EyeNet++ in TensorFlow.
2. Pre-trained models for easy reproducibility.
3. Training and evaluation scripts.
4. Documentation to guide users through setup and usage.

Citation

If you find this work helpful, please cite: paper under review