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
2.61k stars 303 forks source link

Use the RT-DETR on a Jetson device #178

Open KyriakosChris opened 10 months ago

KyriakosChris commented 10 months ago

Currently, I am using the YOLOv7-tiny model on a Jetson device, which exhibits remarkable speed with an inference time of less than 20 milliseconds at a resolution of 640x640. However, when I attempted to employ the rtdetr_r18vd model, the inference time increased approximately threefold, reaching around 60 milliseconds, but the accuracy of the model increased a lot especially on small objects. My goal is to find an RT-DETR model with an inference time of less than 20 milliseconds, ideally without significantly sacrificing resolution to avoid a substantial drop in model accuracy. I am particularly interested in exploring the availability of a smaller RT-DETR model, such as rtdetr_r18vd-tiny, that could potentially meet these specific requirements. Do you have any idea if there is a way to find a smaller RT-DETR model?

Thank you a lot

lyuwenyu commented 10 months ago
  1. Reduce input image resolution (eg. 640 -> 512)
  2. Lightweight backbone (eg. mobilenet)
  3. Less decoder layers
  4. etc

I will release the rtdetrv2 version soon, which will provide a better deployment experience and higher accuracy for small models

Ryanshuai commented 4 months ago

Hi, @lyuwenyu, I feel pity for not seeing you in the CVPR. Any updates on the rtdetrv2? Been waiting for a long time.