We evaluated three types of multi-class learning strategies in a hierarchical compositional framework, namely independent, joint, and sequential training. We conclude that: 1.) Joint and sequential training strategies exert sublinear growth in vocabulary size. 2.) Training time was worst for joint training, while training time even reduced with each additional class during sequential training. 3.) Different training orders of classes did perform somewhat differently. 4.) Training independently resulted in best detection rates, but the discrepancy with the other two strategies was low.
COBISS.SI-ID: 7460180
We developed and evaluated a stochastic optimization approach to learning a compact hierarchical shape vocabulary for object class detection. The optimization iterates between the bottom-up and top-down learning stages, optimally revising the individual layers. We have evaluated the approach on 11 diverse object classes and demonstrated the advantages in terms of speed of inference and detection performance over the previous non-iterated approach as well as the current state-of-the-art methods.
COBISS.SI-ID: 7461716