Loading...
Please wait, while we are loading the content...
Similar Documents
Object recognition by ranking figure-ground hypotheses.
| Content Provider | CiteSeerX |
|---|---|
| Author | Carreira, Joao Li, Fuxin Sminchisescu, Cristian |
| Abstract | We focus on designing methods that can segment images into a number of objects and label each image pixel with the object class to which it belongs (people, animals, bottles, etc). We pursue a segmentation-based front-end to infer plausible regions of support for feature extraction. Differently from existing approaches that dominantly use hierarchical segmentations to fully tile an image with multiple uniform regions [1, 2]- ‘superpixels ’- we compute a set of binary (figure/ground) partitionings, obtained using constrained parametric max flow procedures. This creates a flat family of figure/ground solutions at multiple scales, later assembled into a complete image interpretation using consistency rules. Our segment hypotheses are non-uniform, large in scope, and in practice overlap with entire image objects, or their major parts. This allows reasoning over a larger context and adds the possibility of exploiting Gestalt cues in order to discard implausible segments, rerank the remaining ones, and build rich representations for processing by the next-level recognition module. Our visual object-class recognition is based on continuous value ranking, based on estimates of spatial overlap of each segment hypothesis with putative classes. We differ from existing approaches not only in our assumption that good object segments can be obtained in a feed-forward fashion, |
| File Format | |
| Access Restriction | Open |
| Subject Keyword | Figure-ground Hypothesis Object Recognition Segment Hypothesis Feature Extraction Image Pixel Spatial Overlap Multiple Uniform Region Segmentation-based Front-end Good Object Segment Flat Family Consistency Rule Feed-forward Fashion Entire Image Object Figure Ground Hierarchical Segmentation Implausible Segment Object Class Practice Overlap Putative Class Visual Object-class Recognition Major Part Plausible Region Figure Ground Solution Rich Representation Continuous Value Ranking Segment Image Gestalt Cue Next-level Recognition Module Parametric Max Flow Procedure Multiple Scale Complete Image Interpretation |
| Content Type | Text |
