kornia/rt_detr

Pretrained weights for RT-DETR (Real-Time Detection Transformer),

used by kornia.models.RTDETRDetector.

RT-DETR replaces traditional NMS post-processing with a transformer decoder and an

efficient hybrid encoder combining a multi-scale convolutional backbone with an

intra-scale interaction module. Weights converted from PaddleDetection. CVPR 2024.

Original repo: lyuwenyu/RT-DETR

Weights

| File | Backbone | COCO AP |

|------|----------|---------|

| rtdetr_r18vd_dec3_6x_coco_from_paddle.pth | ResNet-18D | 46.4 |

| rtdetr_r34vd_dec4_6x_coco_from_paddle.pth | ResNet-34D | 48.9 |

| rtdetr_r50vd_m_6x_coco_from_paddle.pth | ResNet-50D-M | 51.3 |

| rtdetr_r50vd_6x_coco_from_paddle.pth | ResNet-50D | 53.1 |

| rtdetr_r101vd_6x_coco_from_paddle.pth | ResNet-101D | 54.3 |

Citation

@inproceedings{zhao2024rtdetr,
    title     = {{DETRs} Beat {YOLOs} on Real-time Object Detection},
    author    = {Zhao, Yian and Lv, Wenyu and Xu, Shangliang
                 and Wei, Jinman and Wang, Guanzhong and Dang, Qingqing
                 and Liu, Yi and Chen, Jie},
    booktitle = {CVPR},
    year      = {2024}
}