Supplementary Material: Deep 360 Pilot: Learning a Deep Agent for Piloting through 360◦ Sports Videos

r(lt(i), l gt t ) = { 1− ‖lt(i)−l gt t ‖2 η , if ‖lt(i)− l gt t ‖2 <= η −1, otherwise (1) where η equals the distance from the center of a viewing angle to the corner of its corresponding NFoV, i.e., √ 32.752 + 24.562 = 40.9 if we define NFOV as spanning a horizontal angle of 65.5◦ with a 4 : 3 aspect ratio. When lt == l gt t , the reward is 1, which is the maximum reward. When ‖lt(i) − l t ‖2 > η, i.e., the predicted viewing angle is not covered by ground truth viewing angle’s NFoV, the reward is -1.