productivity-checker

This project uses a model trained in jetson-inference to measure productivity levels when I'm working. Since I get distracted very easily when I should be doing work, I thought this project would be both a great learning experience and a way to solve a real world problem that I encounter on a daily basis.

Overview

This project uses a resnet-18 model which was trained on 3 classes, sleeping, looking at phone, and working. The dataset used a custom script located in script.py that takes multiple pictures on a camera feed, and saves them into a specified directory. The model is trained on this data and productivity levels are measured based on the confidence of the model's inference.

Classification Model

Model Summary

Model Metrics

Set up

1. SSH into orin
2. Install necessary packages
   sudo apt-get update
sudo apt-get install -y git cmake build-essential libpython3-dev python3-pip python3-numpy
3. Clone jetson-inference
   git clone --recursive --depth=1 https://github.com/dusty-nv/jetson-inference
4. Update submodules
   git submodule update --init --recursive
5. Remove npymath because of a compatibility issue
   sed -i 's/ npymath//g' python/bindings/CMakeLists.txt
sed -i 's/ npymath//g' utils/python/bindings/CMakeLists.txt
6. Install cmake
   mkdir build
cd build
cmake ../

Clone the repository: git clone git@github.com:kezhang29/productivity-checkers.git

Run the project: python3 model.py

Next Steps

• The validation set resembles the training set too much, causing the steep jump in validation accuracy shown in tensorboard. Retraining with a more diverse dataset would improve model results.
• Add more labels