Benchmark evaluation of object segmentation proposal
Hasan, Irtiza (2015-08-28)
Hasan, Irtiza
I. Hasan
28.08.2015
© 2015 Irtiza Hasan. Tämä Kohde on tekijänoikeuden ja/tai lähioikeuksien suojaama. Voit käyttää Kohdetta käyttöösi sovellettavan tekijänoikeutta ja lähioikeuksia koskevan lainsäädännön sallimilla tavoilla. Muunlaista käyttöä varten tarvitset oikeudenhaltijoiden luvan.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:oulu-201508291934
https://urn.fi/URN:NBN:fi:oulu-201508291934
Tiivistelmä
In this research, we provide an in depth analysis and evaluation of four recent segmentation proposals algorithms on PASCAL VOC benchmark. The principal goal of this study is to investigate these object detection proposal methods in an un-biased evaluation framework.
Despite having a widespread application, the strengths and weaknesses of different segmentation proposal methods with respect to each other are mostly not completely clear in the previous works. This thesis provides additional insights to the segmentation proposal methods. In order to evaluate the quality of proposals we plot the recall as a function of average number of regions per image. PASCAL VOC 2012 Object categories, where the methodologies show high performance and instances where these algorithms suffer low recall is also discussed in this work. Experimental evaluation reveals that, despite being different in the operational nature, generally all segmentation proposal methods share similar strengths and weaknesses. The analysis also show how one could select a proposal generation method based on object attributes.
Finally we show that, improvement in recall can be obtained by merging the proposals of different algorithms together. Experimental evaluation shows that this merging approach outperforms individual algorithms both in terms of precision and recall.
Despite having a widespread application, the strengths and weaknesses of different segmentation proposal methods with respect to each other are mostly not completely clear in the previous works. This thesis provides additional insights to the segmentation proposal methods. In order to evaluate the quality of proposals we plot the recall as a function of average number of regions per image. PASCAL VOC 2012 Object categories, where the methodologies show high performance and instances where these algorithms suffer low recall is also discussed in this work. Experimental evaluation reveals that, despite being different in the operational nature, generally all segmentation proposal methods share similar strengths and weaknesses. The analysis also show how one could select a proposal generation method based on object attributes.
Finally we show that, improvement in recall can be obtained by merging the proposals of different algorithms together. Experimental evaluation shows that this merging approach outperforms individual algorithms both in terms of precision and recall.
Kokoelmat
- Avoin saatavuus [31941]