Benchmark: Object Detection for Maritime Search and Rescue

We propose an object detection system for maritime search and rescue as a benchmark problem for verification of neural networks (VNN). The model to be verified is a YOLO (You Only Look Once) deep neural network for object detection and classification and has a very high number of learnable parameters (millions). We describe the workflow for defining and generating robustness properties in the regions of interest of the images, i.e., in the neighborhood of the objects to be detected by the neural network. This benchmark can be used to assess the applicability and the scalability of existing VNN tools for perception systems based on deep learning. URL. Benchmark materials, such as trained models (.onnx), examples of properties (.vnnlib), test images, and property generation procedures, are available at https://github.com/loonwerks/vnncomp2023 .