FOMO_nicla_balloon

A NN-based purple/green balloon detection demo based on the FOMO (faster objects, more objects) architecture from Edge Impulse. The NN is composed of 1 MobileNetV2 block and 2 CNN layers. The input image size to the MobileNet is set to 96 x 96. The network is successfully deployed on a Nicla Vision on QVGA resolution (shrunk to 96x96 before feeding to the NN) with a performance of ~15fps.

Here is a demo video.

ei_object_detection.py

• run this script to inspect the neural network-based detection in a real environment with a real camera
• change LENS_TYPE accordingly

niclavisionsettings.py

• run this script to collect training data at the resolution of QVGA, in the format of a MJPEG video. Extract the individual frames using ffmpeg command ffmpeg -i mjpegvideo.avi -vcodec copy frame%d.jpg.
• Change the variable on line 20 num_frames = 100 to increase/decrease the number of frames to collect.