Parallel Enhanced DETR: Improving Feature Utilization in Object Detection

Deformable DETR(DEtection TRansformer) has improved object detection performance. However, it still suffers from insufficient utilization of features extracted by convolutional networks. To address this issue, this paper proposes the PE-DETR (Parallel Enhanced DETR), which introduces two key innovations. One is the Parallel Encoder, which consists of two branches: the encoder branch leverages the global modeling capability of the Transformer to learn contextual information, while the convolutional branch enhances local feature extraction through convolutional operations. This design effectively integrates global and local features, ensuring the full utilization of the extracted information. Another is the Cross-Scale Feature Enhancement (CSFE) module, which further improves detection performance by fusing multi-scale features. Experimental results show that PE-DETR achieves performance improvements on the COCO 2017 dataset. Compared to Deformable DETR-DC5, integrating the parallel encoder increases the average precision (AP) by 0.9 % and improves the detection accuracy of small objects by 1.1 %, while reducing the number of parameters by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 M}$</tex>. When the CSFE module is removed from the parallel encoder, the AP increases by 0.5 %, and the detection accuracy for small objects improves by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1. 5 \%}$</tex>, along with a reduction of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3 M}$</tex> parameters. In comparison to Conditional DETR, PE-DETR achieves 1.3 % improvement in AP while reducing the number of parameters by 5 M.