Thursday October 28 1:00 PM – Thursday October 28 1:30 PM in Talks I

Image classification in retail: Lessons from the real world

Valentina Bono, Paul Klinger

Prior knowledge:
Previous knowledge expected
Basic knowledge of the ml/deep learning workflow, this talk will mostly empirical and light on technical details

Summary

Deep learning, and specifically convolutional neural networks, are now standard techniques for tackling computer vision tasks. Their results on benchmarks are impressive, but moving from curated standard datasets to a messy real-world setting still presents challenges. We explore some of the challenges encountered when deploying a grocery products classifier to a live in-store environment.

Description

Deep learning, and specifically convolutional neural networks, are now standard techniques for tackling computer vision tasks. Their results on benchmarks are impressive, but moving from curated standard datasets to a messy real-world setting still presents challenges.

In this talk we explore some of the challenges encountered when deploying a simple model to a live in-store environment, specifically for the task of classifying grocery products.

The talk will cover: * Difficulties obtaining & preprocessing data, especially when dealing with legacy data sources. * Unanticipated situations encountered in real world data. * Label quality (both from automatic systems and due to differences in interpretation with human labelers) * Dealing with highly imbalanced datasets. * Distribution shift when deploying models to new (even very similar) settings

We will finish with some future perspectives, looking at the general problem of classifying the whole range of 10s of thousands of products sold in a large store.