FasterRCCN Coco takes 5s each foward pass

Hi there,

Mostly as said in the title, but I am using a FasterRCNN model for object detection on my CPU. Unfortunately, it takes roughly 5s+ for each image to be detected. Is there a way to speed this up? Could it be that I’m using my CPU?


Hi @David_McDonough

Actually Faster-RCNN (and the RCNN family model) is a quite slow detection model, which is a trade off for higher accuracy.
For instance on the GPU (let’s say an NVIDIA GTX 1080) would give you only ~6-8 img/s, which is really slow compare to faster models as YOLOv3 or SSD (detections models yielding slightly lower accuracy).

Could it be that I’m using my CPU?

Anyway, as you suggested, running the model on the GPU would still greatly help.

Thanks Spanev, that makes sense. I’ll try and run it on the GPU. Just wondering, is it possible to use Intel (kaby lake) graphics?

Thanks again

MXNet uses CUDA for GPU acceleration, which is only compatible with NVIDIA GPUs.
This issue about the support of OpenCL exists, this would allow users to use non-NVIDIA GPUs. But the performance is not as good as NVIDIA CUDA and it would require huge efforts to support it.