introduce
In recent years , real-time object detection has been dominated by the YOLO series, and the recently released YOLOv7 (July 6, 2022) is the latest version.
Unlike the current mainstream target detectors that mainly focus on optimizing the architecture, YOLOv7 also focuses on optimizing the training process. The author highlights some optimization modules and optimization methods. This may cost us training with higher accuracy, but not inference cost! !
They call this method a trainable bee bag . They have introduced several of them;
Their two trainable free bee-bag methods, namely reparameterization modules and handling dynamic label assignment to different output layers, have contributed to the development of object detection. We may see reparameterized modules replacing original ones in future object detectors;
YOLOv7 has two architectures. YOLOv7 p5 and YOLOv7 p6 . p6 is larger than p5 . To scale YOLOv7 like concatenation-based models , the authors introduce a new scaling method called dilated and compound scaling . Efficient use of parameters and calculations;
</