Sparse Matching on High-Resolution Maps for Small Object Detection

The effectiveness and precision of most advanced generic object detectors drastically deteriorates when tasked with detecting small objects, as they contain sparse or distorted information. Currently, both anchor-based methods and anchor-free methods are sub-optimal for identifying small targets. It's natural to leverage high-resolution images and low-level feature maps to filter high-quality anchor boxes or anchor points to make predictions. However, computational costs increase in quadratic order due to the large amount of anchors which scales with the dimensions of the employed feature maps. To reduce the amount of computation, we introduce a sparse processing mechanism that integrates seamlessly with YOLOv8, balancing inference speed, accuracy, and resource consumption. To avoid impairing the detection performance of normal targets, we extend YOLOv8's original anchor point assignment strategy by introducing a sparse branch process on the P2 layer specifically focused on small object detection. The sparse branch takes a sparse probability map as input which is generated from the higher-level layer outputs whose values indicate the likelihood of presence of small objects. As a result, we pick these anchor points with high probability as positive samples and build several conv 1×1 modules as SOD heads. This approach ensures that our model attends to both large and small objects, maintaining a high recall for small objects in real-world deployment scenarios. Our experiment exhibits a significant improvement in mean average precision (MAP) by 70% on the SODA-D dataset and high-speed inference on high-resolution feature maps.