Small Underwater Objects 3D Point Cloud (SOUP) Dataset

This repository provides the object recognition benchmark code and label files used to evaluate the SUOP dataset (Zenodo DOI: 10.5281/zenodo.18475883).

Overview

• Purpose: Provide the exact code + labels used for benchmarking object detection/recognition on the SUOP dataset.

• Included:

  • PointNet++: 3D point cloud-based object recognition code.
  • YOLOv8 (Ultralytics): 2D image-based detection code and the corresponding bounding-box label (.txt) files used for training.

• Not included:

  • Raw dataset files and generated images are not distributed here. Images can be reproduced using the same generation pipeline used during benchmarking (e.g., png_make.py) if the SUOP dataset is available locally.

What this repository is for

This repository is intended to help others reproduce the benchmark results (training/inference pipeline) on the SUOP dataset using the provided implementation and label

📁 Dataset Structure

object_detection/
├── PointNet++/
│   └── object_detection_code/
│       ├── model.py
│       ├── dataset.py
│       ├── train.py
│       └── object_detection.py
├── YOLOv8/
│   ├── object_detection_code/
│   │   ├── data.yaml
│   │   ├── ply_change.py
│   │   ├── png_make.py
│   │   ├── train.py
│   │   └── object_detection.py
│   └── bbox_labels/
│       ├── chair/
│       │   ├── chair_range_3m/ (case_XXX.txt ...)
│       │   ├── chair_range_6m/ (case_XXX.txt ...)
│       │   └── chair_range_10m/ (case_XXX.txt ...)
│       ├── drum/ ...
│       ├── dummy/ ...
│       ├── net/ ...
│       └── tire/ ...