STTSBI: A Fast Inference Framework for Small Object Detection in Ultra-High-Resolution Images

Ultra-high-resolution images (UHR images), such as satellite images, aerial photographs, and panoramic images, provide high-quality detection information that enhances small and distant object detection performance. However, UHR images often contain hundreds of millions of pixels, leading to considerable inference time with existing frameworks. To address the aforementioned issue, we extend the YOLOv8 network and propose a novel tiling-based object detection framework called Selective Tiling and Two-Stage Batch Inference (STTSBI). In this framework, the input image is first divided into tiles. Then the inserted pre-classification (PCLS) module filters out tiles without target objects and a two-stage batch inference process is employed to enhance GPU utilization efficiency by maintaining batch processing throughout the entire workflow, thereby significantly reducing inference time. According to our experimental setup, our proposed model can infer an image with over a hundred million pixels in no more than 0.4 seconds, achieving a 10× to 40× increase in inference speed compared to the existing methods. Furthermore, we apply binary quantization to the backbone module of YOLOv8 based on the PhoneBit framework, which further reduces the inference consumption. The binary quantization model achieves a 2× speedup compared to the fullprecision model.