I’ve a few questions relating to the ZED X One GS, which I’d like to use as I require a wider baseline than is available with the other cameras.
Is there a supported GMSL2 capture card for desktop PCs or can it only be used with the Jetson? (Older posts suggest no, but perhaps that has changed since).
Assuming I need to use a Jetson AGX Orin 64GB, what performance can I expect? I would like to run the pair of cameras at 1920x1200 at 60 FPS for my own object detection model, and also the BodyTracking from the ZED SDK. Ignoring the requirements of my own model, will it be fast enough for BodyTracking at that rate? If not, can I do something like run the BodyTracker at a lower frame rate/resolution, while running my own model at the full rate/res? I can’t see anything in the SDK to support this. What rough GPU utilisation can I expect? (My 4090 shows about 35% when running Zed 2i Body Tracking example at 1080/30, but has 8x tensor cores).
This is one of the most frequent questions that we receive, but the reply is still the same: it’s not possible to correctly connect one of the ZED X or ZED X One to a standard, even if provided with a GMSL2 capture card.
This generation of GMSL2 cameras requires NVIDIA Jetson modules as the host because of the absence of an onboard ISP to control camera settings and synchronize stereo and inertial sensor data.
This condition probably will change in the future for new GMSL2 products that we will announce soon. Unfortunately, I cannot reveal more information about this.
Normally an AGX Orin device allows you to retrieve this type of information at a high rate, but configuration tuning is required.
If the performance are not fast enough for your requirements, you could consider using a ZED Box Orin NX 16 GB to acquire two ZED X cameras and leverage the Local Streaming module of the ZED SDK to send the camera streams to a powerful PC on the same network to process the data at higher rates.
Normally an AGX Orin device allows you to retrieve this type of information at a high rate, but configuration tuning is required.
Thanks, but are you able to give any specific figures for an AGX (or NG)? On the blog I see FPS for object detection, but can’t find anything for body tracking.
In the example app source, I see on Jetson depth mode is set to PERFORMANCE rather than ULTRA and the detection model is FAST rather than ACCURATE, which is fine.