def forward_2d_losses(self, out_var, target_var):
out_xy = out_var.narrow(-1, 0, 2)
target_xy = target_var.narrow(-1, 0, 2)
losses = (
euclidean_losses(out_xy, target_xy) +
self._calculate_pixelwise_loss(self.xy_heatmaps[-1], target_xy))
return losses
def forward_2d_losses(self, out_var, target_var):
target_xy = target_var.narrow(-1, 0, 2)
losses = 0
for xy_hm, zy_hm, xz_hm in zip(self.xy_heatmaps, self.zy_heatmaps, self.xz_heatmaps):
losses += self._calculate_pixelwise_loss(xy_hm, target_xy)
losses += euclidean_losses(out_var.narrow(-1, 0, 2), target_xy)
return losses
def forward_2d_losses(self, out_var, target_var):
target_xy = target_var.narrow(-1, 0, 2)
losses = 0
for xy_hm, zy_hm, xz_hm in zip(self.xy_heatmaps, self.zy_heatmaps,
self.xz_heatmaps):
# Pixelwise heatmap loss.
losses += self._calculate_pixelwise_loss(xy_hm, target_xy)
# Calculated coordinate loss.
actual_xy = self.heatmaps_to_coords(xy_hm, zy_hm, xz_hm)[..., :2]
losses += euclidean_losses(actual_xy, target_xy)
return losses
def forward_3d_losses(self, out_var, target_var):
target_xyz = target_var.narrow(-1, 0, 3)
losses = 0
target_xy = target_xyz.narrow(-1, 0, 2)
target_zy = torch.cat([target_xyz.narrow(-1, 2, 1), target_xyz.narrow(-1, 1, 1)], -1)
target_xz = torch.cat([target_xyz.narrow(-1, 0, 1), target_xyz.narrow(-1, 2, 1)], -1)
for xy_hm, zy_hm, xz_hm in zip(self.xy_heatmaps, self.zy_heatmaps, self.xz_heatmaps):
losses += self._calculate_pixelwise_loss(xy_hm, target_xy)
losses += self._calculate_pixelwise_loss(zy_hm, target_zy)
losses += self._calculate_pixelwise_loss(xz_hm, target_xz)
losses += euclidean_losses(out_var, target_xyz)
return losses
def forward_3d_losses(self, out_var, target_var):
out_xyz = out_var.narrow(-1, 0, 3)
target_xyz = target_var.narrow(-1, 0, 3)
target_xy = target_xyz.narrow(-1, 0, 2)
target_zy = torch.cat(
[target_xyz.narrow(-1, 2, 1),
target_xyz.narrow(-1, 1, 1)], -1)
target_xz = torch.cat(
[target_xyz.narrow(-1, 0, 1),
target_xyz.narrow(-1, 2, 1)], -1)
losses = (
euclidean_losses(out_xyz, target_xyz) +
self._calculate_pixelwise_loss(self.xy_heatmaps[-1], target_xy) +
self._calculate_pixelwise_loss(self.zy_heatmaps[-1], target_zy) +
self._calculate_pixelwise_loss(self.xz_heatmaps[-1], target_xz))
return losses
def forward_3d_losses(self, out_var, target_var):
target_xyz = target_var.narrow(-1, 0, 3)
losses = 0
target_xy = target_xyz.narrow(-1, 0, 2)
target_zy = torch.cat(
[target_xyz.narrow(-1, 2, 1),
target_xyz.narrow(-1, 1, 1)], -1)
target_xz = torch.cat(
[target_xyz.narrow(-1, 0, 1),
target_xyz.narrow(-1, 2, 1)], -1)
for xy_hm, zy_hm, xz_hm in zip(self.xy_heatmaps, self.zy_heatmaps,
self.xz_heatmaps):
# Pixelwise heatmap loss.
losses += self._calculate_pixelwise_loss(xy_hm, target_xy)
losses += self._calculate_pixelwise_loss(zy_hm, target_zy)
losses += self._calculate_pixelwise_loss(xz_hm, target_xz)
# Calculated coordinate loss.
actual_xyz = self.heatmaps_to_coords(xy_hm, zy_hm, xz_hm)
losses += euclidean_losses(actual_xyz, target_xyz)
return losses
def mpjpe(actual, expected, included_joints=None):
dists = euclidean_losses(actual, expected)
if included_joints is not None:
dists = dists.gather(-1, torch.LongTensor(included_joints))
return dists.mean().item()
def pck(actual, expected, included_joints=None, threshold=150):
dists = euclidean_losses(actual, expected)
if included_joints is not None:
dists = dists.gather(-1, torch.LongTensor(included_joints))
return (dists < threshold).double().mean().item()