1 / 19

Generating object segmentation proposals using global and local search

Generating object segmentation proposals using global and local search. CVPR2014 Poster. Outline. Introduction Method Experiments Conclusion. Introduction. The sliding window technique suffers from the problem of high computational cost when the number of object categories is large.

jael-horne
Download Presentation

Generating object segmentation proposals using global and local search

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Generating object segmentation proposals using global and local search CVPR2014 Poster

  2. Outline Introduction Method Experiments Conclusion

  3. Introduction The sliding window technique suffers from the problem of high computational cost when the number of object categories is large. we propose a fast method for producing object segmentation proposals by grouping superpixels.

  4. Introduction

  5. Method Superpixels and feature extraction Refined superpixels Local search Global search

  6. Superpixels and feature extraction • We segment the input image into superpixels using two approaches. • The first approach is referred as SLIC and it produces relatively compact superpixelsthat have approximately equal size. • The Second approach is referred as FH and it produces very diverse set of superpixelsthat can be anything from half of the image to a narrow object boundary.

  7. Superpixels and feature extraction we use SIFT descriptors computed on a dense regular grid and RGB values extracted from each pixel. Both descriptors are quantized using visual vocabulary that is learned using training data.

  8. Refined superpixels We first compute a similarity score for each pair of adjacent superpixels. This score is defined for superpixel pair as

  9. Local search the search is split into several parallel branches. This branch is referred as local search, since it considers only superpixel pairs when deciding the next proposal. This approach fails to detect large non-homogeneous objects that consist of diverse set of superpixels.

  10. Global search define the general form of the energy function as

  11. Experiments Experiment 1 Experiment 2 Experiment with other datasets Comparison of execution times

  12. Experiment 1

  13. Experiment 2

  14. Experiment with other datasets

  15. Comparison of execution times

  16. Experiments

  17. Experiments

  18. Conclusion We have presented a fast approach for generating high-quality class-independent object segmentation proposals for color images. Our experimental evaluation with annotated Pascal VOC images shows that the generated region proposals provide accurate segmentations for various kinds of objects. Our approach is approximately as fast as the fastest available comparison method but provides substantially more accurate segmentations.

More Related