Minimal Hough Forest training for pattern detection

This paper assesses Hough Forest configuration parameters with respect to their impact on runtime performance and precision of pattern detection, without large-scale training. The Hough Forest is trained using a very small training set of data and parameters are tuned in a number of experiments, assessing the impact on pattern detection accuracy. A novel method to improve training performance and precision by adaptive selection of the patch size and calculating the used number of patches is introduced. Results are presented using challenging street-scene videos, and demonstrate that the proposed method improves precision and performance over general-purpose Hough Forest parameters used in the literature.