Thesis address: https://arxiv.org/abs/2108.06152
Source code address: https://github.com/Atten4Vis/ConditionalDETR

1. Abstract

The original DETR uses the Encoder and Decoder structures in the Transformer for target detection, and obtains considerable results.In this paper, In order to solve the problem of slow training speed of DETR, a conditional cross-attention mechanism is introduced to achieve faster DETR training.The coss-ateention module in the original DETR relies heavily on content embeddings to achieve the prediction of 4 corners, increasing the high-quality c