-
This repository is to track and store all our experimental AI endeavours, models training, and wishlists.
-
The Robotoff repo is the place to integrate them into production, and file more trivial issues.
-
Most trained Models and useful datasets are attached to releases of this project or releases on robotoff-models.
A Google spreadsheet also tracks active models.
Here are different experiments.
- Nutrition table detection and extraction (2018 GSoc work by Sagar) - integrated in Robotoff, used for the detection part by the Graphnet and TableNet models
- Nutrition Table Extraction (2020 by Sadok, Yichen and Ramzi) - on Graphnet and TableNet
- Basic nutrition extraction for text tables, already in the Robotoff API
-
deployed
-
not deployed:
- EM Lyon Category prediction (2020) - not yet evaluated and integrated
- Category from OCR prediction, Laure (Laurel16) (2021) - not yet evaluated and integrated - Categories maybe too general
-
on-going project @ openfoodfacts/off-category-classification#2
- Labels and Logo detection (Data 4 Good, by Raphael, Charlotte and Antoine - code is duplicated and integrated in Robotoff
- logo-ann (related to logos and labels) - classification using approximate KNN search - deployed in robotoff-ann
- Updating the pre-weighted model to recent publications offers a nice no-effort boost
- Spellcheck (by Wauplin) - code is duplicated and integrated in Robotoff
- ocr-cleaning (please add a description)
- object-detection (related to logos and labels)
You can fork this repository and start your own experiments or use a distinct repository. Please use a AGPL or more permissive but compatible license.
Do not hesitate to join us on #robotoff channel (or #computervision for work relating on images). We will be happy to help you get data, insights and other useful tips.
-
Get the data to start playing with food (see also datasets in this project releases)
- You can see many great analysis of Open Food Facts data in notebooks on Kaggle