Sparse R-CNN: End-to-End Object Detection With Learnable ...

more complex designs to further improve its performance, such as stacked encoder layers [3] and deformable convo-lution network [5], on which a recent work Deformable-DETR [49] is built. However, we align the setting with Faster R-CNN [30, 22] to show the simplicity and effec-tiveness of our method.

  Performance, Testing, Deformable