Abstract

The demise of “segmentation-then-recognition” strategy led to a paradigm shift toward feature-based discriminative recognition with significant success. However, increased complexity in multi-class datasets reveals that local low-level features may not be sufficiently discrimina- tive, requiring the construction and use of more complex structural fea- tures which are necessarily category independent. The paper proposes a bottom-up procedure for generating fragment features which are in- tended to be object part hypotheses. Suggesting that the demise of seg- mentation to generate a representation suitable for recognition was due to prematurely committing to a grouping option in the face of ambi- guities, the proposed framework considers and tracks multiple alternate grouping options. This approach is made tractable by (i) using a me- dial fragment representation which allows for the simultaneous use of multiple cues, (ii) a set of transforms to effect grouping operations, (iii) a containment graph representation which avoids duplicate considera- tion of possibilities, and the estimation of the likelihood of a grouping sequence to retain only plausible groupings. The resulting hypotheses are evaluated intrinsically by measuring their ability to represent ob- jects with a few fragments. They are also evaluated by comparison to algorithms which aim to generate full ob ject segments, with results that match or exceed the state of art, thus demonstrating the suitability of the proposed mid-level representation.