Machine Learning for Object Labelling
Indrajit Saha, Somnath Rakshit, Tanay Ghosh · 2018
Identification of objects from an image using machine learning is an emerging research topic. It is known to the scientific community that the labelling of objects by human intelligence in the current scenario of high volume data consumes a large amount of time and produces an error of 5% as observed in the case of ImageNet. In this regard, machine learning methods have become more popular and effective as they produce better results. However, the performance of different machine learning methods vary while identifying objects from images. To address this fact, an object labelling problem has been formulated by considering the images that contain objects of 10 different classes either in the form of solid, hollow and mixed. In order to label the objects in such three types of images, two different approaches are followed. First, a training dataset is created of size 240 samples where 24 samples are present in each of the 10 classes. Here each class contains equal number of solid and hollow objects. In the second approach, entropy is used to reduce the number of samples in the training dataset needed to achieve similar performance as obtained in the first approach. Thereafter, both the training datasets are applied separately to the seven well-known machine learning methods and one classical method to obtain the final results on three test images. The performance of the methods is demonstrated in terms of accuracy as well as by providing annotated objects in images. A software named ObLab2018 has also been developed to annotate the objects from images. The software and the datasets are provided online at http://www.nitttrkol.ac.in/indrajit/projects/ObLab2018/.