Lifelong_Object_Recognition

Introduction

Code for Intelligent Replay Sampling for Lifelong Object Recognition, IROS 2019. Link to model

Continual learning (CL) is the ability of a model to learn continually from a stream of data, building on what was learnt previously, hence exhibiting positive transfer, as well as being able to remember previously seen tasks.

This task focused on new instance continual learning where number of classes across the tasks will be samw but the domain of the data will change such as occlusion, clutter etc. To know the rules in detail visit this link

There are various methods such as dynamic architectures, regulaization based methods and replay based methods. We propose replay based method.

Dataset used is OpenLoris Dataset. Please refer to competition website for getting the dataset.

Requirements

• Pytorch 1.1
• Python 3.5
• numpy
• pandas
• PIL

Method

Given below is the summary of the proposed algorithm. For details please refer to abstract

For running the code please follow the steps in README.txt

Results

We achieved an mean accuracy of 97.01% at the end of 12th task.

Future Work

This method can be combined with other methods to further reduce the replay size such as synaptic intelligence. We can also think in the direction of latent space replay or psuedo image generations using GANs.