novioleo / mini-amazon-go

Computer Vision Hackathon: Mini Amazon Go

mini-amazon-go

Computer Vision Hackathon: Mini Amazon Go

Setup

Hardware

• Raspberry Pi 4

• camera module

• touch screen + protection case

AWS Kinesis Video Stream on Raspberry Pi

Tutorial: https://docs.aws.amazon.com/kinesisvideostreams/latest/dg/producersdk-cpp-rpi.html

Notes

• The default region is us-west-2. Change to nearest region if applicable.
• Raspbian OS can use apt-get to install shared libs to accelerate compilation of Kinesis Producer C++ SDK (https://github.com/awslabs/amazon-kinesis-video-streams-producer-sdk-cpp/blob/master/install-instructions-linux.md#install-steps-for-ubuntu-17x-and-raspbian-stretch-using-apt-get)
• Setup IAM role for Kinesis Video Stream, and paste the secret key to raspberry
• There's a sample app to upload videos to KVS. ./kinesis_video_gstreamer_sample_app -w <width> -h <height> -f <framerate> -b <bitrateInKBPS> Stream Name (Note that not all resolution works, 640x480 works on Raspberry Pi 4 for me), or you can use src/start_upload.sh.

Trouble shooting

There's potential fix for large latency: https://docs.aws.amazon.com/kinesisvideostreams/latest/dg/troubleshooting.html#troubleshooting-general

SageMaker Notebook

Access Permission

For SageMaker notebook to access KVS, it must have permission AmazonKinesisVideoStreamsFullAccess added to the SageMaker Execution Role when creating the notebook instance

Example notebook for processing KVS

See VideoStream Example.

Example of Training and Test

See src/training_object_detector_yolo3.ipynb and src/test_object_detector_yolo3.ipynb (Yolo3, Recommended)

Or see src/training_object_detector.ipynb and src/test_object_detector.ipynb (SSD)

Or see Beijing CV hackathon