lyuwenyu / RT-DETR

[CVPR 2024] Official RT-DETR (RTDETR paddle pytorch), Real-Time DEtection TRansformer, DETRs Beat YOLOs on Real-time Object Detection. 🔥 🔥 🔥
Apache License 2.0
2.31k stars 258 forks source link

Regnet & DLA34 Backbone Support #454

Closed khadijairfan2345 closed 4 days ago

khadijairfan2345 commented 4 days ago

We have added RegNet, and DLA34 backbone support to the already existing rtdetr_pytorch architecture. These two backbones offer a compelling balance between parameter efficiency and precision. With only 38M and 34M parameters respectively, the rtdetr_dla34 achieves an impressive APval of 49.6 and the rtdetr_regnet surpasses it with 51.6, outperforming existing architectures such as rtdetr_r50vd_m (42M params, APval 51.3). These enhancements not only enhance model performance but also ensure faster inference speeds, making them ideal for real-world applications. I have detailed the relevant information below.

Model Zoo

Model Dataset Input Size APval AP50val #Params(M) FPS checkpoint
rtdetr_regnet COCO 640 51.6 69.6 38 67 url*
rtdetr_dla34 COCO 640 49.6 67.4 34 83 url*

Training

RegNet

1. python tools/train.py -c configs/rtdetr/rtdetr_regnet_6x_coco.yml

DLA34

`2. python tools/train.py -c configs/rtdetr/rtdetr_dla34_6x_coco.yml

lyuwenyu commented 4 days ago

Can you update above Model Zoo into rtdetr_pytorch's readme.md and add an item in homepage 🚀 Updates news

khadijairfan2345 commented 4 days ago

Updated Model Zoo

khadijairfan2345 commented 4 days ago

Added Updates just now, please merge again

lyuwenyu commented 4 days ago

Sorry I merged this pr just now. You can pull latest and push a new pr for Updates info.

khadijairfan2345 commented 4 days ago

Done, I opened another one