Deformable DETR

model paper
SenseTime
Landmark object detection model introducing deformable attention mechanisms to the DETR framework, dramatically improving convergence speed and reducing computational cost. Accepted as oral at ICLR 2021. Multiple pretrained variants released on HuggingFace under the SenseTime organization. Apache-2.0 licensed.
Paper (arXiv)GitHubHuggingFace

Outputs 2

Deformable DETR Model

model

Variants

Name Parameters Notes
deformable-detr — —
deformable-detr-single-scale — —
deformable-detr-single-scale-dc5 — —
deformable-detr-with-box-refine — —
deformable-detr-with-box-refine-two-stage — —

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

paper

arXiv: 2010.04159

Venue: ICLR 2021 (Oral)

visiontransformeropen-source