ShabeenaBanu Hello. We have bought Luxonis camera and performing certain experiments. We are loading a handtracker model/Person detection model on 1080P which is providing FPS of 2 which is too low for real time processing. But for 300x300 resolution the same model outputs fps between 32-35. Even as part of experiments we tried to map the detections(bounding boxes to full frame) which works but the fps is still low,as low as 3fps. How can we use 1080p or higher resolution for real time use cases? 300x300 or other low resolutions wont serve our purpose.
erik Hi ShabeenaBanu , What you have experienced is about the computation limit of the soc (RVC2), some additional info can be found here. In general, this is an edge AI device, and its performance corresponds to that (1.4TOPS for AI). Thoughts? Thanks, Erik
ShabeenaBanu Hi erik Thank you for your reply. Luxonis camera is edge AI Device with Depth capability. But the usefulness of it lies in processing it with higher resolution with depth information. But Unfortunately We will be receiving only 5FPS for 1080P resolution with AI model(I went through the links you shared with me). However, is it possible to process depth information and achieve upto 20-30 FPS for 1080P resolution with AI model on an external device or processor rather than doing it on edge device? Thanks, Shabeena
erik Hi ShabeenaBanu , Yes, you could do that, many have used OAK together with eg. Jetson device which can offer a lot of AI computation, so object detection at 1080P would be possible - best to check their tests though. Thoughts? Thanks, Erik
ShabeenaBanu Hi erik Thanks for your reply. Does it mean we can achieve FPS of 20-30 on Nvidia Jetson with 1080P along with Depth computation? For real time as well as Stored videos? Regards
erik Hi ShabeenaBanu , Nvidia offers different devices with different AI computational power. I think the high end ones should be able to do that, but it would be best to check yourself, as I am not that familiar with jetson series. Thanks, Erik
erik Hi AdamB , Unfortunately, RVC3 doesn't offer any considerable FPS boost compared to RVC2 (RVC2 AI performance docs here). We are now looking at RVC4 (ETA Q1/Q2 next year), and that should have 10x AI performance (at least for yolo models) compared to RVC2. Thoughts? Thanks, Erik
erik SamiUddin we are still evaluating multiple SoCs. In general, we publicly release SN when we release the product itself.
SamiUddin erik That's the great thing. I was also thinking that why LUXONIS only considering movidius. Anyway Best of luck for next series of OAK.
erik SamiUddin We always were considering others as well, but at the time of starting the company (early 2019, see here) it was the best option that we found🙂 Thank you!
