Multi-Head Training Boosting: A Training Method to Increase the Accuracy of Low-Parameter Models in Multiple Object Tracking

Multiple object tracking (MOT) is a problem with widespread applications in real-world tracking scenarios. In this context, object detection plays a crucial role in the performance of MOT. However, there are significant challenges in applying object detection models in industry, such as the trade-off between speed and accuracy. To address this issue, multiple versions of models are typically created with varying numbers of parameters depending on the intended use. For example, in the YOLO family, there are different versions like YOLO-n, YOLO-s, YOLO-m, etc. Models with fewer parameters offer faster processing speeds suitable for real-time tasks, but at the expense of lower accuracy. To optimize this trade-off, we propose a method can enhance the accuracy of low-parameter models without compromising computational speed. This is based on the fact that low-parameter models often saturate quickly during training, leading to limited knowledge acquisition from the dataset. The proposed method can be easily applied to both one-stage and two-stage object detection models. This approach promises significant benefits in terms of cost, resource efficiency, and scalability of various models in the future. Based on experiments, when applied to ByteTrack, this method increases the MOTA of the Bytetrack model with backbone YOLOX-s from 79.7 to 83.8.