How can one feed 4 camera feeds directly into snpe without having it read from the input_lists.txt list of raw files? Can one not avoid having to read and write the images to files and instead capture them with OpenCV and directly feed them as cv::Mat objects into the snpe object and ML model?
Thanks,
Hi 22imonreal
I have not understand your qestions clearly , you have 4 camera streamings which will be used for SNPE?
What do you mean one feed? In general , each of streaming could be directly inputed to SNPE without file reading/writing.
Please provide more information about your use case.
Thanks
Kevin
You can start 2 threads, one is for camera streaming capture and preocess, the streaming image will stored in buffer which will be used for another process for SNPE inference. The SNPE inference thread will do SNPE processing.
How do you think about it.
Hello and thanks for the prompt reply.
Yes, I will have 4 cameras and will have to combine the frames from 2 cameras at each timestamp into a single 640, 320, 6 tensor that should then be inputed into the model directly.
How do I capture the camera feed into a buffer, then process the images, then recombine the frames into a tensor in another buffer and finally feed this last buffer into the model using snpe without the need to read and write files using different processes as you suggested?
(The only examples I have seen take the inputs from an input_list.txt (see snpe-net-run and snpe-sample applications provided by Qualcomm))
Thanks.
You may use sample steps for SNPE in
https://developer.qualcomm.com/forum/qdn-forums/hardware/qualcomm-roboti...