TPH-YOLOv5++: Boosting Object Detection on Drone-Captured Scenarios with Cross-Layer Asymmetric Transformer

Object detection in drone-captured images is a popular task recent years. As drones always navigate at different altitudes, the object scale varies considerably, which burdens optimization of models. Moreover, high-speed and low-altitude flight cause motion blur on densely packed objects, leads to great challenges. To solve two issues mentioned above, based YOLOv5, we add an additional prediction head detect tiny-scale objects replace CNN-based heads with transformer (TPH), constructing TPH-YOLOv5 model. TPH-YOLOv5++ proposed significantly reduce computational cost improve speed TPH-YOLOv5. In TPH-YOLOv5++, cross-layer asymmetric (CA-Trans) designed while maintain knowledge this head. By using sparse local attention (SLA) module, information between other can be captured efficiently, enriching features heads. VisDrone Challenge 2021, won 4th place achieved well-matched results 1st model (AP 39.43%). Based CA-Trans further increase efficiency achieving comparable better results.