Project Description
Project Identification
Project name: Image outlining and tagging by unsupervised refinement kernel
Project Title (Romanian): Extragerea si etichetarea regiunilor de interes din imagini prin rafinare nesupervizata la nivel de pixel
Contract No. 22PED/2017
Program name within PN III: Funding Application for experimental demonstration projects (Programme 2, Subprogramme 2.1 - "Competitivitate prin
cercetare, dezvoltare si inovare – Proiect experimental - demonstrativ")
Funding Agency: Executive Unit for Higher Education,
Research, Development and Innovation Funding
(UEFISCDI)
Project code: PN-III-P2-2.1-PED-2016-0292
Abstract
During the past years, there has been an increased interest in the development of machine learning algorithms, focused on applications such as scene understanding, contextual segmentation, object detection, object recognition and classification. A common prerequisite for these methods is the existence of large annotated databases which can be used as both training and test data, targeted on the application at hand.
The existence of such datasets is generally limited to a small number of classes and scenarios, based on constraints imposed by its designated use. Changing marking modes or adding new objects and functionalities to such large image repositories implies both extensive human effort and the availability of extended periods of time for completion. Moreover, available datasets are either too small, unbalanced or very limited.
In order to overcome these limitations, we propose a project which implements a new solution for pixel-level semi-automatic annotation of large datasets, without constraints based on application or number of identified classes. Our aim is to develop a very reliable marking and querying workflow which addressed the problems related to big data paradigms.
Through these results, we aim to increase the actual technology readiness level (TRL) from a value of 2 at the beginning of the project, to a value of 4 at its completion. We argue that the proposed toolchain showcases increased efficiency when compared to other existing methods, with the proof of effectiveness being the usage of all available information in a single image (pixel-wise labeling), thus providing training samples compatible with most of the existing computer vision algorithms.
Another confirmation of efficacy is given by a deep knowledge of the database through the advanced statistical information provided by the proposed workflow, as better knowledge of available data represents the primary condition for developing successful machine learning (ML) algorithms.
The project coordinator is University Politehnica from Bucuresti. The project will run for a period of 18 months, with a budget of 600,000 lei.