Get our free extension to see links to code for papers anywhere online!

Chrome logo Add to Chrome

Firefox logo Add to Firefox

DA-DETR: Domain Adaptive Detection Transformer by Hybrid Attention

Mar 31, 2021
Jingyi Zhang, Jiaxing Huang, Zhipeng Luo, Gongjie Zhang, Shijian Lu

The prevalent approach in domain adaptive object detection adopts a two-stage architecture (Faster R-CNN) that involves a number of hyper-parameters and hand-crafted designs such as anchors, region pooling, non-maximum suppression, etc. Such architecture makes it very complicated while adopting certain existing domain adaptation methods with different ways of feature alignment. In this work, we adopt a one-stage detector and design DA-DETR, a simple yet effective domain adaptive object detection network that performs inter-domain alignment with a single discriminator. DA-DETR introduces a hybrid attention module that explicitly pinpoints the hard-aligned features for simple yet effective alignment across domains. It greatly simplifies traditional domain adaptation pipelines by eliminating sophisticated routines that involve multiple adversarial learning frameworks with different types of features. Despite its simplicity, extensive experiments show that DA-DETR demonstrates superior accuracy as compared with highly-optimized state-of-the-art approaches.

Share this with someone who'll enjoy it:

   Access Paper Source

Share this with someone who'll enjoy it: