def chamfer_distance(pred, gt, sqrt=False):
"""
Computes average chamfer distance prediction and groundtruth
:param pred: Prediction: B x N x 3
:param gt: ground truth: B x M x 3
:return:
"""
if isinstance(pred, np.ndarray):
pred = Variable(torch.from_numpy(pred.astype(np.float32))).cuda()
if isinstance(gt, np.ndarray):
gt = Variable(torch.from_numpy(gt.astype(np.float32))).cuda()
pred = torch.unsqueeze(pred, 1)
gt = torch.unsqueeze(gt, 2)
diff = pred - gt
diff = torch.sum(diff**2, 3)
if sqrt:
diff = guard_sqrt(diff)
cd = torch.mean(torch.min(diff, 1)[0], 1) + torch.mean(
torch.min(diff, 2)[0], 1)
cd = torch.mean(cd) / 2.0
return cd
def chamfer_distance_one_side(pred, gt, side=1):
"""
Computes average chamfer distance prediction and groundtruth
but is one sided
:param pred: Prediction: B x N x 3
:param gt: ground truth: B x M x 3
:return:
"""
if isinstance(pred, np.ndarray):
pred = Variable(torch.from_numpy(pred.astype(np.float32))).cuda()
if isinstance(gt, np.ndarray):
gt = Variable(torch.from_numpy(gt.astype(np.float32))).cuda()
pred = torch.unsqueeze(pred, 1)
gt = torch.unsqueeze(gt, 2)
diff = pred - gt
diff = torch.sum(diff**2, 3)
if side == 0:
cd = torch.mean(torch.min(diff, 1)[0], 1)
elif side == 1:
cd = torch.mean(torch.min(diff, 2)[0], 1)
cd = torch.mean(cd)
return cd
def to_variable(tensor, device='cpu'):
if isinstance(tensor, np.ndarray):
if device == 'cpu':
tensor = Variable(torch.from_numpy(tensor.astype('float32')))
elif device == 'gpu':
tensor = Variable(
torch.from_numpy(tensor.astype('float32')).cuda())
return tensor
def chamfer_distance_single_shape(pred,
gt,
one_side=False,
sqrt=False,
reduce=True):
"""
Computes average chamfer distance prediction and groundtruth
:param pred: Prediction: B x N x 3
:param gt: ground truth: B x M x 3
:return:
"""
if isinstance(pred, np.ndarray):
pred = Variable(torch.from_numpy(pred.astype(np.float32))).cuda()
if isinstance(gt, np.ndarray):
gt = Variable(torch.from_numpy(gt.astype(np.float32))).cuda()
pred = torch.unsqueeze(pred, 0)
gt = torch.unsqueeze(gt, 1)
diff = pred - gt
diff = torch.sum(diff**2, 2)
if sqrt:
diff = guard_sqrt(diff)
if one_side:
cd = torch.min(diff, 1)[0]
if reduce:
cd = torch.mean(cd, 0)
else:
cd1 = torch.min(diff, 0)[0]
cd2 = torch.min(diff, 1)[0]
if reduce:
cd1 = torch.mean(cd1)
cd2 = torch.mean(cd2)
cd = (cd1 + cd2) / 2.0
return cd
