Continual learning for object detection

Abstract

This Master Thesis is contextualized in the context of the need to count containersin different cities in Spain for garbage collection. In order to perform this countingautomatically, to save time and resources, it is proposed to use the technique of objectdetection on images captured from a car.At the research proposal level, the hypothesis is as follows: ”it is possible to createa model/system that takes advantage of the knowledge acquired from the cubes of acity to facilitate learning in new cities, without significantly worsening its results inthe previously known city”. To achieve this, one has to deal with a problem deriveddirectly from the training algorithm of a neural network, the Catastrophic Forgetting.This problem implies that, once a model with certain weights is trained to solve a newtask (e.g., recognition of a new class) its results will be worse on the previously knowntask/class.In order to try to solve this problem, a current called Continual Learning (alsoknown as Incremental Learning or Lifelong Learning) arises within the area of Arti-ficial Intelligence, which introduces techniques based on reproduction, regularizationor isolation of parameters, among others, to try to mitigate the effect of CatastrophicForgetting in the results of the model, retaining the knowledge prior to the retrainingof the model with the new data.Specifically, in this project we have worked to achieve a system that is capable ofdetecting containers of new cities without losing accuracy in the detection of previouslyknown cities. To this end, we have carried out annotation tasks, adaptation of a GitHubrepository where the EfficientDet model is used for object detection, measurement andmitigation of catastrophic forgetting and finally creation of an interface to facilitatethe use of the application to a potential user.