Detail of Publication
| Text Language | Japanese |
|---|---|
| Authors | Koichi KISE, Kazuto NOGUCHI, and Masakazu IWAMURA |
| Title | Improving Efficiency and Robustness of Specific Object Recognition by Increasing Reference Feature Vectors |
| Journal | Proceedings of MIRU2009 |
| Presentation number | OS12-1 |
| Pages | pp.335-342 |
| Reviewed or not | Not reviewed |
| Month & Year | July 2009 |
| Abstract | This paper concerns specific object recognition based on local features. In general, increasing the number of local features for indexing (reference feature vectors) by generative learning enables us to improve the recognition rate. In this paper, we show that generative learning is also effective for shorten the processing time. This paradoxical effect (i.e., shorter time with more reference feature vectors) is achieved by cascading recognizers. Generated local features allow us to terminate the recognition process earlier. In other words, generative learning is effective to reduce the search space for finding nearest neighbors. From the experimental results using 10,000 reference images, 6.6 times reference feature vectors enabled us both to reduce the processing time to 2/3 from the original, and to improve the recognition rate by 12.2\%. Another experiment with 1 million reference images indexed by 2.6 billion reference feature vectors yielded the recognition rate of 90\% in 59ms/query. |
- Following file is available.
- Entry for BibTeX
@InCollection{KISE2009, author = {Koichi KISE and Kazuto NOGUCHI and Masakazu IWAMURA}, title = {Improving Efficiency and Robustness of Specific Object Recognition by Increasing Reference Feature Vectors}, booktitle = {Proceedings of MIRU2009}, year = 2009, month = jul, presenID = {OS12-1}, pages = {335--342} }