Deeply Optimized Hough Transform: Application to Action Segmentation

Hough-like methods (Implicite Shape Model, Hough forest, 9 ...) have been successfully applied in multiple computer vision fields like 10 object detection, tracking, skeleton extraction or human action detection. 11 However, these methods are known to generate false positives. To handle 12 this issue, several works like Max-Margin Hough Transform (MMHT) or 13 Implicit Shape Kernel (ISK) have reported significant performance im14 provements by adding discriminative parameters to the generative ones 15 introduced by the Implicit Shape Model (ISM). In this paper, we pro16 pose to use only discriminative parameters that are globally optimized 17 according to all the variables of the Hough transform. To this end, we 18 abstract the common vote process of all Hough methods into linear equa19 tions, leading to a training formulation that can be solved using linear 20 programming solvers. 21 Our new Hough Transform significantly outperforms the previous ones 22 on HoneyBee and TUM datasets, two public databases of action and 23 behaviour segmentation. 24