In loss_crit.py, there are some comments,
def forward(self, pred_3D_lanes, gt_3D_lanes, pred_hcam=None, gt_hcam=None, pred_pitch=None, gt_pitch=None): """ :param pred_3D_lanes: predicted tensor with size N x (ipm_w/8) x 3*(2*K+1) :param gt_3D_lanes: ground-truth tensor with size N x (ipm_w/8) x 3*(2*K+1) :param pred_pitch: predicted pitch with size N :param gt_pitch: ground-truth pitch with size N :param pred_hcam: predicted camera height with size N :param gt_hcam: ground-truth camera height with size N :return: """
I have some questions on detailed loss design:
- What's the meaning of ipm_w/8 ?
- what's the meaning of k?
- Why groud truth dimensions are like this?
Is there any detailed inforamtion to refer? I found no information on this kind in the paper
In loss_crit.py, there are some comments,
def forward(self, pred_3D_lanes, gt_3D_lanes, pred_hcam=None, gt_hcam=None, pred_pitch=None, gt_pitch=None): """ :param pred_3D_lanes: predicted tensor with size N x (ipm_w/8) x 3*(2*K+1) :param gt_3D_lanes: ground-truth tensor with size N x (ipm_w/8) x 3*(2*K+1) :param pred_pitch: predicted pitch with size N :param gt_pitch: ground-truth pitch with size N :param pred_hcam: predicted camera height with size N :param gt_hcam: ground-truth camera height with size N :return: """I have some questions on detailed loss design:
Is there any detailed inforamtion to refer? I found no information on this kind in the paper