Comparison of YOLO model performance across mobile devices of different classes

This dataset contains measurements of frames per second (FPS) and inference time for multiple generations of YOLO models deployed in TensorFlow Lite format. The experiments were conducted on four mobile devices of different performance classes: Infinix GT 30 Pro Poco X5 Pro Samsung A34 Ulefone Armor 14 Pro All models were evaluated using a fixed input resolution of 320×320 pixels. Inference was performed in real-time conditions using a custom Android application. The application utilizes 4 CPU cores, and input frames are obtained from the device camera, converted from the camera YUV format to ARGB_8888 Bitmap, and resized prior to inference. Notes and limitations:The use of a fixed number of CPU cores may affect performance scalability across devices.Additional preprocessing (color conversion and resizing) contributes to the overall inference time.Measurements were obtained under real-time conditions and may vary depending on thermal and system load factors.A value of -1 in the dataset indicates that the model failed to execute correctly, either due to runtime errors, unsupported operations, or invalid output (e.g., random bounding boxes). Funding:This research was conducted within the framework of the research project “Navigation system for the visually impaired” (state registration No. 0126U002460), funded by the Ministry of Education and Science of Ukraine.