원문정보
초록
영어
In this paper, we present a novel object detection scheme using only local contour fragments. A sample fragment extraction method decomposes a whole contour into several parts. Then, the candidate locations of corresponding fragments in test images are detected by a modified Hausdorff distance with punishment on clutter edge regions. The most probable locations are selected by Multiclass Discriminative Field (MDF). Finally, contours of the objects can be drawn with these locations and sample fragments. Our major contributions are simplifying the MDF by an undirected graph constructed by the candidate locations and directly selecting the fragment locations by this MDF. The results on our postmark database and the ETHZ database from internet show that the proposed scheme is practicable.
목차
1. Introduction
2. Local Fragment Detection
2.1. Sample Fragment Extraction
2.2. Candidate Location Detection
3. Fragment Selection by MDF
3.1. MDF Introduction
3.2. Parameter Estimation
3.3. Model Inference
4. Experiments
5. Conclusion
Acknowledgements
References