Selective Search as Object Recogntion
This is the outline of my robotics 417 graphics project. This
project was required work for the
Introduction to Robotics
taught by Gregory Dudek.
This project was submitted by
Hannan Syed. This webpage provides a brief outline of results and conclusion drawn in the report found here.
A zip of the code, images, and rendered images (after running the code -- used in the evaluation) can be found here.
Abstract
This paper evaluates the selective search algorithm implemented by J.R.R. Uijlings et al.. The selective search algorithm addresses the problem of object recognition. In particular the selective search has emphasis on the inherit hierarchical structure of images. This is done by combining segmentation for object recognition with exhaustive search. The advantage of exhaustive search is that is aims to capture all object locations, and the advantage of segmentation is that it uses image structure to guide the search for object locations. The selective search results in a small set of data-driven, class-independent, high quality locations. The results of selective search have been outstanding with exceptional scores across the Pascal Image challenges. This paper evaluates external potential challenges where the algorithm may fail to recognize an object. These instances may include camouflaged object, which may be obvious to a human but not so much to the selective search algorithm.
Results
The image set used in the evaluation can be found here:
Some of the tests run are illustrated below.
Chameleon Base Case:
Some successful attempts at recognizing the object in the image can be seen on the right by the blacking out of the background.
Chameleon Test Case:
This image on the other hand shows absolutely zero indication that an object was recognized as the original image is returned.
Conclusions
"...The toughest test for the algorithm was working with camouflage images. This is certainly understandable as even humans have some trouble recognizing camouflaged insects or people. The first image was a chameleon which was similar to texture and color to its surroundings on the ground of a forest. Generally the algorithms boxes which are printed out attempt to identify the object in the image, however in this case the boxes returned the entire image as seen in figure 4 above. Additional tests were done with camouflaged people, some obvious to the human eye and other not so obvious. None of the these images fared well at all. The entire image was returned without any indication of recognition attempt. These images posed the greatest challenge for the algorithm because of the challenge to differentiate between texture and color."
Acknowledgements
I would like to acknowledge Uijlings, J. R.R. et al. for their work on Selective Search For Object Recognition, which was the basis for this paper.