Object Segmentation by Graph Partitioning ∗

Segmentation and recognition have long been treated as two separated processes. We propose a mechanism based on spectral graph partitioning that readily couple the two processes into one. A part-based object recognition system detects parts, supplies their partial segmentations and an evaluation of how well possible labellings on these parts go together based on the statistics of object spatial configurations. We integrate the top-level part grouping into the low-level pixel grouping based on feature similarity to segregate objects of interest from the background. We demonstrate that object part consistency and image pixel coherence can work together to eliminate local false positives and labelling ambiguities at high level, while at the same time overcome occlusion and weak contours at low level.