Some questions about paper

[JoJoliking]

As claiming in the paper, the proposed methods tracks the object in occlusion and uses a new datasets for videos training and propagation. So I have two questions about this:

1. Do this methods lead to higher FP ? Because the paper have not shown this metric.
2. I notices that crowded human datasets is used in model training. So, compared the proposed synthetic data, which data can bring much gains? Thank you for your outstanding works. I am interesting in this paper and want to know more about the details. I am looking forward to getting your responds.

[pvtokmakov]

Hi,

thanks for you interest in our work!

1. We avoid false positives by additionally training a box visibility classifier, and only outputting visible detections at test time.
2. CrowdHuman dataset increases visible person detection accuracy, but training on synthetic data allows us to learn to localize people behind occlusions. These two factors are complementary, and both necessary for achieving top results.