vapaau/TUB

The Turbid Underwater Baseline (TUB) dataset addresses the lack of curated data for vision research in highly turbid underwater environments. It supports tasks such as underwater image enhancement and restoration with ground truth, object detection and segmentation under severe visibility degradation, and multi-view reconstruction.

TUB consists of 33 static scenes captured inside a cylindrical water tank using four synchronized GoPro11 cameras. Scene geometry, camera poses, and object arrangements remain fixed while turbidity is systematically increased. Turbidity is induced by adding measured amounts of oat milk and quantified using nephelometer readings, ensuring reliable and reproducible ground truth.

Each scene is recorded at 10 turbidity levels, later regrouped into low (0–10 NTU), medium (10–20 NTU), and high (20+ NTU) turbidity categories to provide clear visual separation and balanced data splits. The dataset contains 1,315 images total.

All scenes are fully annotated at the instance-segmentation level under clear-water conditions, producing over 16,000 masks. These annotations are propagated to all turbidity levels, enabling high-confidence ground truth even when objects are no longer visible to human annotators.

TUB is currently the largest underwater dataset of its kind providing paired clear and extreme-turbidity imagery with reliable annotations.

Citation: