Tao Hu

Tao Hu

Auto-generated profile, based on publicly available information

Tao Hu

Wuhan Institute of Technology, China

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
24 May 2024
Verified
Conference Paper
Open access
JWJianghao Wei
Jianghao Wei
WHWei Huang
Wei Huang
THTao Hu
Tao Hu
刘强刘 强
刘 强
JYJie Yu
Jie Yu
JHJiahuan Huang
Jiahuan Huang
The recently proposed DEtection TRansformer (DETR) removes hand-designed components and has become a new paradigm for object detection. However, the application of DETR is limited due to its expensive inference cost. While edge devices provide a solution for more challenging scenarios, they are unable to accommodate the expensive inference cost of DETR. To address this issue, we propose a Lightweight, Efficient, and Decoupled DETR, called LED-DETR. We divide DETR into two parts: the feature extractor and the object querier. In the feature extractor, we introduce lightweight convolutional neural networks as the backbone and efficient attention mechanisms as the self-attention mechanism for more lightweight. In the object querier, we incorporate the conditional spatial query for more performant. Compared with the original DETR, LED-DETR achieves a good trade-off between compute, memory footprint, and performance.