#######################################################################
#                                                                     #
# Questions for                                                       #
# "Learning Correspondence from the Cycle-Consistency of Time"        #
# by Xiaolong Wang, Allan Jabri, Alexei A. Efros                      #
#                                                                     #
# Answers due on Wednesday **July 17, 2019 at 10am**                  #
#         to mayern@cs.uni-freiburg.de                                #
#                                                                     #
#######################################################################

1. What is the point of "feature" matching? Why not simply match image patches?
   (1-2 sentences)

2. Which frames can be skipped by the "skip cycle" in Sec. 3.1.2? Can you 
   think of drawbacks to this formulation, e.g. situations in which both the
   "tracking" and the "skip cycle" could fail? 
   (2-3 sentences)

3. Pretraining ResNet-50 on ImageNet seems to outperform the presented method
   everywhere. Do you think this is a weakness of this paper? Why (not)?
   (1-2 sentences)