Market-Computer-Vision

To clone this repository, type git clone --recursive https://github.com/aisu-programming/Market-Computer-Vision.git.

Notice that the option --recursive is necessary because of submodules.

Download models

Type the command sh download-weights.sh to download all model's weights files.

Environment setup

1. Type the command sh setup-torch.sh to install PyTorch (version 1.9) for Raspberry Pi 32-bits OS with Python 3.7.
2. Type the command sh setup-torchvision.sh to install torchvision (version 0.10) for Raspberry Pi 32-bits OS with Python 3.7.
3. Type the command sh setup-others.sh to install all other requirements.
4. Setup the Google Sheet API. See tutorial.

Update this project

Type the command git pull then sh download-weights.sh to get the newest model weight file.

Usage

The main command is python3 detect.py.

There are serveral options can be add:

• --fd: enable the Footfall-Detection function. (default: False)

  Example: python3 detect.py --fd.

• --fd-img-size <int>: the size of input images for the Footfall-Detection task will be resize to this. (default: 64)

  Example: python3 detect.py --fd-img-size 32 will set the size of input images to 32x32.

• --sed: enable the Shelf-Empty-Detection function. (default: False)

  Example: python3 detect.py --sed.

• --sed-img-size <int>: the size of input images for the Shelf-Empty-Detection task will be resize to this. (default: 256)

  Example: python3 detect.py --sed-img-size 128 will set the size of input images to 128x128.

• --sed-alert-amount-thres <float>: the amount threshold to view a stock as empty. (default: 0.5)

  Example: python3 detect.py --sed-alert-amount-thres 0.3 will set the alert amount to 0.3.

• --sed-smoothing-len <int>: the length of the smoothing array which to prevent unstable predictions. (default: 10)

  Example: python3 detect.py --sed-smoothing-len 20 will set the length of the smoothing array to 20.

• --test: use the testing images in test directory to test the inference. (default: False)

  Example: python3 detect.py --test.

• --view-img: show real-time camera images. (default: False)

  Example: python3 detect.py --view-img.

• --sleep <int>: the sleep time between every inference. (default: 0)

  Example: python3 detect.py --sleep 3 for sleeping 3 seconds after every image.

• --save-img: save inferenced images to the output directory or not. (default: False)

  Example: python3 detect.py --save-img.

• --save-img-interval <int>: save inferenced images to output directory per seconds. (default: 900)

  Example: python3 detect.py --save-img-interval 60 for saving image every 1 minute (60 seconds).

• --save-csv: save results to csv file or not. (default: False)

  Example: python3 detect.py --save-csv.

• --save-csv-interval <int>: save results to csv file per seconds. (default: 900)

  Example: python3 detect.py --save-csv-interval 60 for saving records every 1 minute (60 seconds).

• --save-google-sheet: save results to google sheet or not. (default: False)

  Example: python3 detect.py --save-google-sheet.

• --save-google-sheet-interval <int>: save results to google sheet per seconds. (default: 900)

  Example: python3 detect.py --save-google-sheet-interval 60 for saving records every 1 minute (60 seconds).

Combine above options or edit default value of each options in detect.py.

Some data

	Raspberry Pi 3B	Raspberry Pi 4B
Installation time (setup-*.sh + download-weights.sh)	31 minutes	23 minutes
Footfall-Detection speed (input img size = 64)	2~4 sec / per inference	1.4 sec / per inference
Shelf-Empty-Detection speed (input img size = 256)	6~8 sec / per inference	2~3 sec / per inference

About the OS

I used the Raspberry Pi Imager.

The OS I chose is the Raspberry Pi OS (Legacy), which is 32-bits and with the Python version 3.7.