IFCD: Inverted feature extraction for enhancing curling stone detection

The movement data of curling targets is of great significance for the analysis and research of curling. However, in real-life curling competitions, the curling volume is limited and easy to be occluded, and the venue background illumination is complicated. To address these challenges, a curling target detection model, IFCD, based on Inverted Feature Extraction Network (IFNet) is proposed. IFNet allocates more resources to deal with high-resolution features without introducing additional computational burdens, thus avoiding feature loss caused by inappropriate downsampling. Moreover, a Dynamic Feature Fusion module is introduced in the Neck network to suppress background interference and reduce the feature confusion. In addition, the parameter-independent Four-Scale Decoupled Detection Head is introduced to reduce the conflict between classification and regression tasks and enhance the model’s multi-scale adaptability. IFCD achieves a 0.974 mAP@.5 (Mean Average Precision) on Curling, a regular curling dataset, and 0.723 mAP@.5 on Curling_hard, a complex curling dataset with numerous occlusions.