Hey!
I am trying to benchmark the ZED 2 object detection model with other pre-trained models.
I noticed that the model is downloaded and loaded as a .model file. Is this a ZED-specific file extension, such as .svo files for recording? Can I load this model file with using e.g.: tensorflow, or is it like a “black box” for us?
Thank you in advance!
Hi,
Yes the models used by the ZED SDK are specific and proprietary, unfortunately.
I understand this may not be realistic or convenient to label a new dataset but the only solution, for now, would be to capture the data from the ZED2 to be able to benchmark it against others.
Thank you for the fast answer!
Do you provide any performance metrics on the ZED 2 object detection models, or the measurements are not public yet?
Currently, we don’t communicate on that. I’ll let you know if this changes.
Just in case we do in the future, what would be the format of performance metrics you would like?
It would be interesting to see the performance of the ZED 2 object detection model on some benchmark datasets like COCO or PASCAL VOC. And metrics like AP on different IoU levels, precision and recall curves.
Just one more question regarding the obj.detection model. Is there perhaps a way to override an image stream topic with e.g. benchmark pictures from COCO, and have the obj. detection model run on those?
Hi and thank you for your answers!
If providing specific benchmarks is not an option at the moment, could you perhaps give us at least some comparison benchmarks between the 3 models (ACCURATE, MEDIUM, FAST)? Because based on my limited experiments, I don’t seem to find any difference, but later on this will matter a lot.
Knowing a benchmark tradeoff between speed and accuracy between all 3 models would be really helpful! Thanks in advance for your reply!