Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
924
924
End of preview. Expand in Data Studio

TUB-synthetic

Synthetic turbid counterparts of the TUB (Turbid Underwater Baseline) dataset, released with the paper "Bridging the turbidity gap with synthetic data" (Ismiroglou, Pedersen, Bengtson, Moeslund; Pattern Analysis and Applications, 2026).

The paper extends the ICCV 2025 workshop paper Sea-ing Through Scattered Rays. The journal version evaluates the resulting data quantitatively and shows that training on STSR images improves generalization in real turbid scenes.

Repository structure

baseline/ and STSR/ hold the main synthetic datasets. Each comes in two versions, one per monocular depth estimator used to produce z(x):

baseline/                      # reference IFM
β”œβ”€β”€ depthanythingv2/           # depth from Depth Anything V2
└── zoedepth/                  # depth from ZoeDepth
STSR/                          # full STSR pipeline
β”œβ”€β”€ depthanythingv2/
└── zoedepth/
ablations/
β”œβ”€β”€ baseline_standard_depth/   # baseline, fixed depth span between z_min and z_max
β”‚   β”œβ”€β”€ depthanythingv2/
β”‚   └── zoedepth/
β”œβ”€β”€ STSR_standard_depth/       # STSR, fixed depth span between z_min and z_max
β”‚   β”œβ”€β”€ depthanythingv2/
β”‚   └── zoedepth/
β”œβ”€β”€ only_inhomogeneity/        # baseline + GRF inhomogeneity only
β”‚   β”œβ”€β”€ depthanythingv2/
β”‚   └── zoedepth/
β”œβ”€β”€ only_forwardscattering/    # baseline + forward scattering only
β”‚   β”œβ”€β”€ depthanythingv2/
β”‚   └── zoedepth/
β”œβ”€β”€ baseline_no_depth/         # baseline, no depth map
└── STSR_no_depth/             # STSR, no depth map

Per-variant layout

Each leaf folder holds 1,317 samples across 33 scenes (Exp1–Exp33) and 4 cameras (Cam1–Cam4). File stems match across the three subfolders and mirror TUB's Exp<scene>_Cam<view>_<idx> naming:

<variant>/
β”œβ”€β”€ images/          # Exp10_Cam1_001.png  – synthetic turbid image, 1920Γ—1080 RGB
β”œβ”€β”€ labels/          # Exp10_Cam1_001.txt  – instance segmentation labels
└── image_metadata/  # Exp10_Cam1_001.json – IFM parameters used for this image

Labels use the YOLO segmentation format, one object per line:

<class_id> x1 y1 x2 y2 ... xn yn     # polygon vertices normalized to [0, 1]

Metadata stores the parameters used to render each image:

Key Description
a absorption coefficient a(Ξ»), one value per sampled wavelength
b scattering coefficient b(Ξ»)
beta beam attenuation coefficient Ξ²(Ξ») = a(Ξ») + b(Ξ»)
K diffuse downwelling attenuation coefficient K_d(Ξ»)
phi forward-scattering blur scale (Οƒ = Ο† Β· z); null when forward scattering is disabled
g fraction of scattered light leaving the line of sight, used in G = a + g Β· b
mu backlight brightness factor (B∞ = μ · b · E / G)
z_min, z_max metric scene-distance range (m) that the estimated depth map is rescaled to
gamma gamma applied to the relative depth map before rescaling
grf_min, grf_max range of the Gaussian Random Field that models medium inhomogeneity

Citation

If you use this data, please cite:

@article{ismiroglou2026bridging,
  title     = {Bridging the turbidity gap with synthetic data},
  author    = {Ismiroglou, Vasiliki and Pedersen, Malte and Bengtson, Stefan H. and Moeslund, Thomas B.},
  journal   = {Pattern Analysis and Applications},
  volume    = {29},
  number    = {4},
  pages     = {186},
  year      = {2026},
  publisher = {Springer},
  doi       = {10.1007/s10044-026-01770-4}
}
Downloads last month
316

Collection including vapaau/TUB_synthetic