We have tried the official object_detection_sample_ssd (shipped along with openVINO R4), and compared the perf between NCS (Myriad2) and NCS2 (Myriad X), we found the perf gain is very limited. Say for NCS (Myriad2) is about 5.8 fps, and for NCS2 (MyriadX) is about 6.3 fps. Does anyone know if any settings can be tuned? Since NCS2 is said to have about at least 3x perf on this network.
Any hints will be highly appreciated~
Hello, anyone help to answer this question please? Since I did see intel show ssd-mobilenet demo with ncs2, which was able to run at around 30fps, on the intel AIDC event.
Where does your ssd-mobilenet model comes from? I tried the mobilenet-ssd downloaded by using model_downloader. That network is trained by using caffe and its input resolution is 300x300. Its performance on ncs2 is expected on my env, about 30 fps when running with 'benchmark_app' in the package.
The model of your own training, the computing power is similar to mobilenet_ssd, but the input is 512*512. Is NCS2 particularly sensitive to the size of the model input?
Hi, I have a problem. Using SSD Mobilenet v1 with 90 classes_num in config I get 0.05 seconds inference time, however , when I use my own trained model with only 1 classes_num I get 4.7 seconds inference time. Have I missed something. I use ssd_v2_support.json for model optimization.