Distributed Information Systems Laboratory LSIR

Human-in-the-Loop Image Classification

Project Details

Human-in-the-Loop Image Classification

Laboratory : LSIR Semester / Master Proposal




Description:
General Overview

Our project aims at accelerating the process of building AI classifiers for image classification. In traditional human-in-the-loop approaches, the most time-consuming and expensive phase is spent in the tedious work of labelling datasets. We propose to dramatically reduce this labelling phase by leverage image similarity techniques. Human time and attention are precious. Instead of presenting the annotators with an image, asking them to label it, we show multiple images in a clustered interface so that similar samples are clustered near each other. Annotators can quickly annotate a few thousand of training samples which is enough to start training a new classifier using transfer learning. Our visualization tool also enables active learning. After a few iterations of training, annotators can visualize what samples the model is confused and then focus the labeling on those. This is done repeatedly as needed by the classifier to increase accuracy and reduce confusion in the long run.

Description of the project

K-means is currently used for clustering. K-means is fast to compute, but the number of clusters k needs to be defined beforehand. In this project, we want to explore *density-based or graph-based clustering algorithm* that can not only cluster the data points but can also determine the number of clusters as well based on the density of the data, such as DBSCAN or Chinese Whispers. Large clusters are not easily visualisable. In a second step, we want to explore methods for choosing the most representative points from clusters created by these clustering algorithms.

Potential starting point: https://www.pyimagesearch.com/2018/07/09/face-clustering-with-python/

Deliverables: codebase with documentation

Prerequisites
  • Familiar with Python
  • Creativity, spirit, initiative and pro-active
  • Knowledge of Linux and related tools
Preferred, but not required
  • Experience in Machine Learning
  • Experience in Computer Vision

Send me your CV: remi.lebret@epfl.ch.


Site:
   
Contact: Rémi Lebret
Useful Links

Project Guidelines