Efficient Regression of General-Activity Human Poses from Depth Images: Supplementary Material

mean average precision using grid search with a step size of 0.05m in the range [0.05, 0.60]m. Table 1 shows that depending on which error metric is used (i.e., does missing an occluded joint count as a false negative or not?), the optimized length thresholds are quite different. In the right column we see that when the model is penalized for missing occluded joints, it makes use of longer range votes to maximize mAP. In some cases, such as head, feet, and ankles, the difference is rather large. Intuitively this makes sense: occluded joints tend to be further away from visible depth pixels than non-occluded joints. This experiment used a forest trained on 30k images.