I’m the process of finding object detection networks that can run 100% on DLA (Orin AGX) with an inference performance on a 900x700 px image of about 100 fps. I tried some of the TAO networks (ssd, retinanet with a resnet34 backbone), but when I use DLA, there’s always a fallback to GPU and the inference performance is very low ~ 8 fps. Using the same network, running on the GPU, the performance is ~ 200 fps
Can someone please help point me to some resources, or if anyone has some experience with similar issue pls share how you went about solving it.
Thank you for the pointer. I understand that these are pretrained models. I tried TAO models with these arch but couldn’t achieve these levels of performance. Seems like these models are different from TAO models…
Could you please help point me to some resources to fine-tune/transfer learning on my own dataset ?