UW-TransStereo: Underwater Stereo Vision Dataset for Transparent Object Detection & Ranging

Robust detection and ranging of transparent debris in water is challenging due to refraction, low contrast, backscatter, and turbidity. We present a paired, multi-environment stereo dataset acquired with a stereo camera, comprising 25 paired stereo vision recordings totalling over 9,000 stereo frame pairs across air and four underwater conditions (freshwater, freshwater with enhancement, saltwater, and saltwater with suspended pellets), covering three bottle targets (two transparent plastic bottles with filled and unfilled variants, and a brown glass medicine bottle), with per-frame detections, depth and dimension (H x W x D) estimates, over 5,000 labelled measurements across environment–bottle combinations (frames, detections, ROI samples), and analysis-ready detector spreadsheets and scripts. To promote reproducible use, we provide a standard evaluation protocol with well-defined accuracy measures—mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R2), and mean signed error (bias)—and report the proportion of frames with valid estimates (``valid%''), including vertical region-of-interest (ROI) sampling (upper/centre/lower). We include a lightweight ``basicEnhance'' variant to illustrate how pre-processing can be benchmarked against the same protocol without retraining models. Users can load the paired .svo2/.xlsx files to benchmark alternative detectors or compare ranging strategies across the five environmental conditions. The release is intended as a neutral testbed for benchmarking underwater transparent-object perception and for stress-testing stereo ranging under refractive media. File naming, storage structure, and missing-value conventions (including explicit use of not-a-number, NaN) are documented to ease integration. All raw stereo streams, detector outputs, qualitative screenshots, and analysis scripts are published here with a persistent digital object identifier (DOI) to enable citation and long-term reuse. The repository contains multiple versions of an object detection system, progressing from simple confidence-based detection to advanced 3D analysis with the following key features: Real-time object detection using YOLO models (YOLOv5/YOLOv8) 3D depth integration with ZED stereo cameras Object dimension measurement (width, height, depth) Multi-point distance sampling for accurate measurements Real-time data export to Excel with timestamps Screenshot capture with visual distance markers Pause/resume functionality for detailed analysis SVO recording support for offline processing 3D visualization with OpenGL rendering