SynSoda Underwater Synthetic Dataset

Overview

SynSoda is a single-class underwater object detection dataset focused exclusively on the “soda can” class. The images are generated with Stable Diffusion 3.5 using prompts provided in soda_can_prompts.txt, targeting photorealistic soda cans in clear euphotic-zone water under different viewpoints, colors, and lighting conditions. [file:2]

The dataset is intended for training and evaluating underwater trash detection models, domain adaptation methods, and robust perception pipelines for ROVs/AUVs.

Source Datasets and Domains

The synthetic images are guided by soda-can instances extracted from several underwater datasets:

• COU (lake, sea, and pool environments)
• TrashCan
• Walia et al.
• UNO underwater dataset

From each of these sources, only the soda-can class is used as a visual reference for the generative process.

Each synthesized image is then transformed to resemble three distinct environmental domains derived from COU:

• Lake-like appearance
• Sea-like appearance
• Pool-like appearance

Color transfer and histogram matching are applied to approximate each environment’s illumination and color distribution.

Data Generation Pipeline

1. Prompt design

   All prompts are stored in soda_can_prompt.txt. The prompts describe photorealistic soda cans underwater in clear euphotic-zone water, with variations in:

   • Can color (e.g., blue, red, green, black, gold)
   • Orientation (centered horizontally vs. vertically)
   • Time of day (daytime vs. nighttime)
   • Minimal background clutter and neutral seabed

2. Base image synthesis

   For each text prompt, Stable Diffusion 3.5 generates one guided base image.

3. Per-prompt multiplicity

   From each guided image, 20 images are produced from the prompts.

4. Domain adaptation / style transfer

   Each of the 20 images is expanded to 4 variants:

   • Original diffused image
   • Lake-matched version (color transfer + histogram matching)
   • Sea-matched version (color transfer + histogram matching)
   • Pool-matched version (color transfer + histogram matching)

   This inflates each original guided image to 80 images in total.

Annotations and Labels

• The dataset contains a single class:

  • 0 → soda_can

• Class definitions and train/val split information are stored in data.yaml.

• Images and labels follow a standard YOLO-style convention:

  • One label file per image
  • Each line in a label file:
    class x_center y_center width height (all coordinates normalized to [0, 1])

Make sure to adjust data.yaml paths to match your local folder structure before training.

Directory Structure

A typical layout for the dataset is:

SynSoda/
  images/
    train/
      *.png
    val/
      *.png
  labels/
    train/
      *.txt
    val/
      *.txt
  data.yaml
  soda_can_prompts.txt
  README.md

• data.yaml: dataset configuration for YOLO-based training (paths, number of classes, class names).
• soda_can_prompts.txt: Stable Diffusion prompts used to generate the base images.

Intended Use Cases

• Underwater trash / debris detection (single-class detection)
• Synthetic-to-real domain adaptation across lake, sea, and pool styles
• Ablation studies on the effect of color transfer and histogram matching
• Robust perception for ROV/AUV control loops where soda cans are target objects

Limitations

• Only one object class (soda_can) is included.
• Backgrounds are relatively simple by design (minimal clutter, neutral seabed), which may not fully represent complex real-world environments. [file:2]
• All base images are synthetic; real-world fine-tuning is recommended for deployment scenarios.

Citation

If you use SynSoda in your research, please cite this dataset:

R. J. Indrayanto, SynSoda: Synthetic Underwater Soda Can Dataset, 2026.
(DOI to be available.)

Attribution

SynSoda was constructed using soda-can objects and environmental characteristics derived from the following datasets:

• COU Underwater Trash Dataset
  DOI: 10.13020/1vwe-2707

• UNO Underwater Dataset
  Project page: https://www.lirmm.fr/uno

• TrashCan Underwater Debris Dataset
  DOI: 10.13020/g1gx-y834

• Walia et al. – UTD2 Dataset
  Dataset page: https://universe.roboflow.com/utd-0dazj/utd2-hyo53