BACHELOR Assignment
Place recognition by image retrieval
Type: Bachelor CS
Period: TBD
Student: Sievers, C.T. (Cas, Student M-CS)
If you are interested please contact :
Introduction:
Visual place recognition is an important task for applications of mapping and localization in robotics, 3D reconstruction, autonomous driving and so on. Recognizing a place from pictures taken at different viewpoints is a difficult task as the geometry of the scene is subject to perspective changes. Further challenges in outdoor environments are given by changes of illumination, day-and-night variations, and changes to the appearance of the scenes due to seasons. One recent application of place recognition is to support robot localization in garden and agricultural scenarios, where the textured appearance of the scene makes the automatic visual analysis task difficult. We recently made publicly available a large data set for place recognition in gardens [1,2].
Assignment:
One possible approach for place recognition is to formulate it as an image retrieval problem, retrieving the images from a database that are most similar to the query images [3]. The assignment is to evaluate the performance of techniques for image retrieval, especially those based on the bag-of-words representations, and compare their performance with those of convolutional networks (e.g. NetVLAD [4]). Thinking of modifications and optimizations to the bag-of-words approach is encouraged.
References:
[1] https://github.com/marialeyva/TB_Places
[2] M. Leyva-Vallina, N. Strisciuglio, M. López Antequera, R. Tylecek, M. Blaich and N. Petkov, "TB-Places: A Data Set for Visual Place Recognition in Garden Environments," in IEEE Access, vol. 7, pp. 52277-52287, 2019.
[3] María Leyva-Vallina, Nicola Strisciuglio, Nicolai Petkov, Place Recognition in Gardens by Learning Visual Representations: Data Set and Benchmark Analysis, CAIP 2019