Small Object Detection in Complex Images: Evaluation of Faster R-CNN and Slicing Aided Hyper Inference

Abstract

Small object detection has many applications, including maritime surveillance, underwater computer vision, agriculture, traffic flow analysis, drone surveying, etc. Object detection has made notable improvements in recent years. Despite these advancements, there is a notable disparity in performance between detecting small and large objects. This gap is because small objects have less information and a weaker ability to express features. This paper investigates the performance of Faster Region-Based Convolutional Neural Networks (R-CNN), one of the most popular and user-friendly object detection models for head detection and counts in artworks rather than images of real humans. The impacts of Slicing Aided Hyper Inference (SAHI) on the enhancement of the model’s capability to detect small heads in large-size images are also being analyzed. The Kaggle-hosted Artistic Head Detection dataset was used to train and evaluate the proposed model. The effectiveness of the proposed methodology was demonstrated by integrating SAHI into two other object detection models, Cascaded R-CNN and Adaptive Training Sample Selection (ATSS). The experimental results reveal that applying SAHI on top of any object detector enhances its ability to recognize and detect tiny and various scaled heads in large-scale images, which is a significant challenge in numerous applications. At a confidence level of 0.8, the SAHIenhanced Faster R-CNN achieved the best private Root Mean Square Error (RMSE) score of 5.31337, while the SAHI-enhanced Cascaded R-CNN obtained the highest public RMSE score of 3.47005.