Reproduce result for Kitti benchmark

[nghiattran]

@MarvinTeichmann How reproduce results in your Multinet paper? I tried to add resnet decoders and compare them with the VGG decoder but when I submit results onto http://www.cvlibs.net/datasets/kitti/, it turns out that VGG can only reach 50% on moderate. I used min_conf=0.50, tau=0.35 and exactly same setting as in provided pretrained VGG.

[MarvinTeichmann]

Strange. What is your validation score?

[nghiattran]

Score produced by tv-analyze:

	train easy	train moderate	train hard	val easy	val moderate	val hard	speed (msec)	speed (fps)	post (msec)
VGG	93.25%	84.10%	69.09%	94.27%	86.32%	70.78%	104.4554	9.5735	3.5804

I don't have a screenshot of Kitti result because I submitted more than a week ago and it has been deleted.

[robin2002]

I also submitted a result using the model (KittiBox_pretrained.zip), which is provided by author (MarvinTeichmann) and got the 45.22% for the moderate in a car detection task. (Easy: 71.99% and Hard: 36.83%) . Have you produced results using the Kitti test dataset?

[nghiattran]

This is the result for the test set. This result is higher than what I originally got but it is very low compared to what is in the official paper.