Few-shot object detection using self-supervised learning

PreciBake, is a company based in Munich, New York and Mumbai, developing AI solutions for the food-tech and baking industry. Our AI team is continuously working on developing and improving our machine learning algorithms for tasks such as image classification, object detection and tracking.

A challenge often faced with real world data sets is that they have uneven data distributions with respect to different properties such as class labels, light conditions, scene environment or camera angles. Due to these data domain shifts a model trained in the lab environment might work poorly when deployed in a new environment where the data “looks” different to what the model has been trained on.

To overcome these issues in practice it is usually necessary to fine-tune models with labeled data from the target environment. But even fine-tuning might require hundreds of labeled examples per class to yield good accuracies. As the number of vision tasks increases it becomes increasingly more expensive to maintain high quality data-sets, especially in environments that are non-stationary.

The goal of this project is to develop an object detection model that can overcome the limitations by 1. utilizing large unlabeled datasets to pretrain for proxy tasks using self-supervised learning, and 2. learn the target task e.g. detecting a specific object by using only a few labeled images and possible unlabeled images from the target environment. Such a model could potentially be trained once for any object detection task without labeling effort and then configured on installation to a specific task and environment with only a few descriptive examples.