TilapiaVisionDataset: Annotated image dataset for oreochromis niloticus

Precision aquaculture requires automated monitoring systems capable of supporting fish management and production control in intensive farming environments. This article describes the TilapiaVisionDataset [3], a curated image dataset developed for object detection and counting of Nile tilapia (Oreochromis niloticus) under controlled aquaculture conditions. The dataset was generated using an experimental water flume designed to simulate fish handling and monitoring scenarios, operating with a controlled water flow rate of 2700 L/h to reproduce realistic movement and interaction patterns while maintaining a low-stress acquisition procedure. Image acquisition was performed using a fixed high-resolution camera positioned above the flume, resulting in 8512 images with a spatial resolution of 1920 × 1080 pixels. A total of 63,004 fish instances were manually annotated using bounding boxes following the YOLO annotation format. The dataset includes scenes characterized by high stocking density, averaging 7.40 fish per frame, and incorporates common visual challenges observed in aquaculture environments, such as fish overlap, partial occlusion, motion variability, reflections, and illumination changes. To support machine learning workflows, the dataset is organized into training (70%), validation (15%), and testing (15%) subsets. The annotation structure and dataset organization enable direct compatibility with modern computer vision frameworks and object detection architectures, including single-stage and two-stage deep learning models. The dataset can be reused for the development, benchmarking, and validation of automated fish detection, counting, and monitoring systems. Potential applications include algorithm training, comparative evaluation of detection models, robustness analysis under dense biological scenarios, and research involving computer vision techniques applied to aquaculture monitoring environments.