Deformable DETR: Deformable Transformers for End-to-End Object Detection

TL;DR

Deformable DETR introduces a modified Transformer architecture with deformable attention that focuses on key sampling points, significantly improving object detection performance and training efficiency, especially for small objects.

Contribution

The paper presents Deformable DETR, a novel Transformer-based model with deformable attention modules that enhance feature processing and reduce training time.

Findings

Outperforms DETR on COCO benchmark

Achieves better small object detection

Requires 10 times fewer training epochs

Abstract

DETR has been recently proposed to eliminate the need for many hand-designed components in object detection while demonstrating good performance. However, it suffers from slow convergence and limited feature spatial resolution, due to the limitation of Transformer attention modules in processing image feature maps. To mitigate these issues, we proposed Deformable DETR, whose attention modules only attend to a small set of key sampling points around a reference. Deformable DETR can achieve better performance than DETR (especially on small objects) with 10 times less training epochs. Extensive experiments on the COCO benchmark demonstrate the effectiveness of our approach. Code is released at https://github.com/fundamentalvision/Deformable-DETR.