Deformable DETR is the official open-source implementation of the influential object detection model of the same name, published on GitHub where it has earned about 4,000 stars. It improves on the original DETR transformer detector with deformable attention modules that converge far faster and perform much better on small objects.
Standard DETR attends to every pixel of a feature map, which makes training slow and small-object detection weak. Deformable DETR instead has each query attend to a small set of sampling points around a reference, and extends the idea across multiple feature scales. The result is a detector that reaches better accuracy with roughly ten times fewer training epochs. The repository, built on the SenseTime Fundamental Vision team’s codebase, includes training and evaluation code for COCO, pretrained weights and configs for several variants, including two-stage and iterative bounding-box refinement versions. The codebase was last updated in 2024.
Deformable DETR is fully free and open source under the Apache 2.0 license. There is no commercial version, hosted service or paid tier; the code and pretrained models can be used in research and products at no cost.
Deformable DETR is for computer vision researchers and engineers who want a strong, well-documented transformer-based object detector to study, reproduce results with, or adapt to their own datasets.