In fcos_mono3d_head.py (FCOS3d head), i notice that the orientation regression target is computed by function add_sin_difference and the loss is smoothL1 loss, which means we would like to minimize $smoothL1(sin(\theta_p, \theta_t))$ where theta_p and theta_t denotes the predict angle and gt angle.In original paper, FCOS divide the local yaw to 2 bins and regress the residual angle w.r.t the "anchor angle" of the corresponding bin, but i think the implementation here is actually directly predict the local yaw instead of the residual angle. Are there something wrong i've got? Many thanks!