基于知识蒸馏的轻量化Transformer目标检测

Object Detection of Lightweight Transformer Based on Knowledge Distillation

在自动驾驶领域,目标检测的高效性和准确性尤为重要,基于Transformer结构的目标检测方法逐渐成为主流,省去了复杂的锚点生成和非极大值抑制。针对现有方法计算成本高和收敛速度慢的问题,设计了一种基于池化操作的轻量化Transformer目标检测模型(LPT),包含了池化主干网络和双池化注意力机制,设计了针对DETR(detection transformer)模型的通用知识蒸馏方法,将预测结果、查询向量和教师提取的特征作为知识传递给轻量化的Transformer模型,帮助其提升精确度性能。通过在MS COCO 2017数据集上的实验,验证经过蒸馏的LPT模型在自动驾驶中的应用潜力,实验结果表明:本文方法具有较好的准确性,与一些先进的方法相比具有一定优势。

In autonomous driving,the efficiency and accuracy of object detection are significant.Object detection based on Transformer structure has gradually become the mainstream method,eliminating the complex anchor generation and non-maximum suppression(NMS).It has problems of high computing cost and slow convergence.An object detection model of the based lightweight pooling transformer(LPT)is designed,which contains a pooling backbone network and dual pooling attention mechanism.A general knowledge distillation method is intended for the DETR(detection transformer)model,which transfers prediction results,query vector,and features extracted by the teacher as knowledge to the LPT model to improve its accuracy.To verify the application potential of the distilled LPT model in autonomous driving,extensive experiments are conducted on the MS COCO 2017 dataset.The results show that the method has great efficiency and accuracy,and is competitive with some advanced techniques.