Stereo Fish Dataset for BézierFusion: 3D Curve Estimation in Aquaculture

Description

This dataset contains binocular stereo images and annotations for underwater fish detection and 3D body length measurement, supporting the research paper "Bézierfusion: A Unified Fish Detection and Measurement Paradigm". Dataset Overview: The dataset consists of two distinct subsets covering different aquaculture scenarios: Dataset 1 (Indoor Environment): Collected from an indoor grouper (Plectropomus leopardus) farm featuring a stable blue tank background. Dataset 2 (Outdoor Environment): Collected from outdoor ponds, featuring complex challenges like weak textures and suspended particles. Note on Data Completeness: This repository contains a representative subset of the full dataset. It includes the complete Test Set and a sample of the Training Set. This subset shares the exact same directory structure and annotation format as the full dataset, allowing researchers to run the provided code and reproduce the evaluation metrics reported in the paper.

Steps to reproduce

1. Data Acquisition: Images were captured using a custom-designed underwater binocular measurement system. The setup captured synchronized top and side views of the fish. 2. Annotation Protocol (YOLO Format): The dataset uses a Bézier curve-based annotation strategy. The annotations are provided in standard YOLO format (.txt files). File Structure: Each image in the images/ folder has a corresponding text file in the labels/ folder. Coordinates: All coordinates are normalized to [0, 1]. Row Format: Each line represents one fish instance: <class_id> <x_center> <y_center> <width> <height> <P1_x> <P1_y> <P2_x> <P2_y> <P3_x> <P3_y> <P4_x> <P4_y> (Where P1 to P4 represent the 4 Bézier control points). 3. Usage: Users can directly load this dataset into the provided codebase to verify the model's detection and curve fitting performance.

Institutions

• Hainan University

Categories

Funders

• Hainan Seed Industry549 Laboratory

  Grant ID: B23H10004

Licence