yactouat / sws_waste-sorter

an image classifier to help with waste triage

👋 I'm Yacine.

I'm a software developer and an optimistic futurist. I am passionate about problem solving, tech, AI, software engineering, and web development.

You can visit my Obsidian publish website to get to know me better!

sws_waste-sorter's Issues

styled flag selector in navbar

build a programmable trashbin that will open only if you put the right item in front of it

allow webcam access to the web app' to take pictures of waste

internationalize the frontend

• flag selector in navbar
• English is the default language
• French is a supported language

auto delete old Google Artifact Registry images

better visual feedback on prediction output

• better contrast ratio
• feedback picture and feedback color should be of same fixed size

...

show projects contributors in the app'

apply the drive train approach to the waste sorter problem

export model on https://huggingface.co/docs/hub/timm when it's performant enough

better handling of "not a waste item"

review data augmentations techniques that are being used in training the model

document the local dev and CI/CD pipeline

make the dataset public

give more context to the users

• project context
• technological content
• social media, GitHub, and notebook links
• GitHub readme
• newsfeed

AI model should be able to recognize different parts of a waste item to (to be thrown in different trash bins)

use image segmentation for this ?

build a programmable trash bin that will dispatch the thrown item within it in the right inner container

waste sorter mobile impl

• port existing web features
• should include camera feed to classify items live

...

choose a `PyTorch` model that combines speed and accuracy better than `resnet18`

see https://www.kaggle.com/code/jhoward/which-image-models-are-best/ for comparison

convnext seems to be a good candidate

allow the model to derive information form packaging writings

this could help the model to compare with packagings where writings are not available

add relevant OpenGraph meta

set up GitHub sponsoring

• implement GitHub sponsors
• UI should contain a CTA for sponsoring

migrate the notebook code to to this repo

set up GitHub projects

rewrite the frontend in Typescript

relevant robots.txt file

list all the issues, with their description, in the website

fancy logo + favicon

turn the frontend into a PWA

train the classifier with parts of the images of the dataset

either use fastai functions or add items to the dataset (parts from the training images)

build a conveyor belt that will dispatch an item to the right container based on its type

select a suitable name and get a domain name for this app'

static landing-page for SEO

full view of the official trash bins/color code documentation on click

put a loader while the uploaded file is being processed by the `/predict` endpoint

CICD pipeline should include building the frontend

train the model to be able to recognize waste items in low luminosity / low res setup

make the Docker image leaner

allow users to send their predictions feedback in order to improve the model

set up GitHub Conversations

prod deployment of Flask app

• Flask debug mode off
• React i18n debug mode off
• browse docs to know what else needs to be done for prod deployment

...

review NPM and dependabot security issues

style the frontend

build a trashbins system that tells the user where to throw items with a light + bip

allow the app' to adapt to other bins/color codes systems throughout the World

@ritik2358 , do you have the same kind of sorting system in place in the city of Bangalore ? I'm thinking of expanding the app's features to the cities of the contributors ;)

solve npm vulnerabilities

benchmark what connections can be made with Citeo app' use-cases

https://www.citeo.com/

QR code on packagings sytems to complement the solution