Detrex is an open-source research platform from IDEA Research that unifies the DETR family of transformer-based vision models in one codebase. Released under the Apache-2.0 license with more than 2,300 GitHub stars and updates through 2025, it collects the many DETR variants that were originally scattered across separate repositories.
Since the original DETR reframed object detection as set prediction with transformers, the family has exploded: Deformable DETR, DAB-DETR, DN-DETR, DINO, Mask DINO and more, each with its own repo and quirks. Detrex re-implements them on a shared foundation built on Detectron2, with common training, evaluation and model-zoo conventions. That makes it practical to compare variants fairly, swap components โ backbones, necks, detection heads โ and reproduce published results on benchmarks like COCO for detection, segmentation and pose estimation from a single, maintained codebase.
Detrex is free and open source under the Apache-2.0 license.
Detrex is for computer vision researchers and engineers working with transformer detectors who want one reliable platform for benchmarking, fine-tuning and extending DETR-style models instead of maintaining a fork per paper.
The detrex model zoo reports COCO results for each variant under standard 12-epoch and longer training schedules, and the platform supports several backbones, including ResNet, Swin Transformer and FocalNet, so detection heads can be compared on equal footing. The project also provides tools for visualizing predictions and analyzing model behavior, plus conversion paths for exporting trained models toward deployment. For a research platform, its documentation and benchmark tables are unusually complete.