AnySplat / src /loss /loss_distill.py
alexnasa's picture
Upload 243 files
2568013 verified
import torch
import torch.nn as nn
import torch.nn.functional as F
from copy import copy, deepcopy
from src.model.encoder.vggt.utils.pose_enc import pose_encoding_to_extri_intri
from src.model.encoder.vggt.utils.rotation import mat_to_quat
from src.utils.point import get_normal_map
def extri_intri_to_pose_encoding(
extrinsics,
intrinsics,
image_size_hw=None, # e.g., (256, 512)
pose_encoding_type="absT_quaR_FoV",
):
"""Convert camera extrinsics and intrinsics to a compact pose encoding.
This function transforms camera parameters into a unified pose encoding format,
which can be used for various downstream tasks like pose prediction or representation.
Args:
extrinsics (torch.Tensor): Camera extrinsic parameters with shape BxSx3x4,
where B is batch size and S is sequence length.
In OpenCV coordinate system (x-right, y-down, z-forward), representing camera from world transformation.
The format is [R|t] where R is a 3x3 rotation matrix and t is a 3x1 translation vector.
intrinsics (torch.Tensor): Camera intrinsic parameters with shape BxSx3x3.
Defined in pixels, with format:
[[fx, 0, cx],
[0, fy, cy],
[0, 0, 1]]
where fx, fy are focal lengths and (cx, cy) is the principal point
image_size_hw (tuple): Tuple of (height, width) of the image in pixels.
Required for computing field of view values. For example: (256, 512).
pose_encoding_type (str): Type of pose encoding to use. Currently only
supports "absT_quaR_FoV" (absolute translation, quaternion rotation, field of view).
Returns:
torch.Tensor: Encoded camera pose parameters with shape BxSx9.
For "absT_quaR_FoV" type, the 9 dimensions are:
- [:3] = absolute translation vector T (3D)
- [3:7] = rotation as quaternion quat (4D)
- [7:] = field of view (2D)
"""
# extrinsics: BxSx3x4
# intrinsics: BxSx3x3
if pose_encoding_type == "absT_quaR_FoV":
R = extrinsics[:, :, :3, :3] # BxSx3x3
T = extrinsics[:, :, :3, 3] # BxSx3
quat = mat_to_quat(R)
# Note the order of h and w here
# H, W = image_size_hw
# fov_h = 2 * torch.atan((H / 2) / intrinsics[..., 1, 1])
# fov_w = 2 * torch.atan((W / 2) / intrinsics[..., 0, 0])
fov_h = 2 * torch.atan(0.5 / intrinsics[..., 1, 1])
fov_w = 2 * torch.atan(0.5 / intrinsics[..., 0, 0])
pose_encoding = torch.cat([T, quat, fov_h[..., None], fov_w[..., None]], dim=-1).float()
else:
raise NotImplementedError
return pose_encoding
def huber_loss(x, y, delta=1.0):
"""Calculate element-wise Huber loss between x and y"""
diff = x - y
abs_diff = diff.abs()
flag = (abs_diff <= delta).to(diff.dtype)
return flag * 0.5 * diff**2 + (1 - flag) * delta * (abs_diff - 0.5 * delta)
class DistillLoss(nn.Module):
def __init__(self, delta=1.0, gamma=0.6, weight_pose=1.0, weight_depth=1.0, weight_normal=1.0):
super().__init__()
self.delta = delta
self.gamma = gamma
self.weight_pose = weight_pose
self.weight_depth = weight_depth
self.weight_normal = weight_normal
def camera_loss_single(self, cur_pred_pose_enc, gt_pose_encoding, loss_type="l1"):
if loss_type == "l1":
loss_T = (cur_pred_pose_enc[..., :3] - gt_pose_encoding[..., :3]).abs()
loss_R = (cur_pred_pose_enc[..., 3:7] - gt_pose_encoding[..., 3:7]).abs()
loss_fl = (cur_pred_pose_enc[..., 7:] - gt_pose_encoding[..., 7:]).abs()
elif loss_type == "l2":
loss_T = (cur_pred_pose_enc[..., :3] - gt_pose_encoding[..., :3]).norm(dim=-1, keepdim=True)
loss_R = (cur_pred_pose_enc[..., 3:7] - gt_pose_encoding[..., 3:7]).norm(dim=-1)
loss_fl = (cur_pred_pose_enc[..., 7:] - gt_pose_encoding[..., 7:]).norm(dim=-1)
elif loss_type == "huber":
loss_T = huber_loss(cur_pred_pose_enc[..., :3], gt_pose_encoding[..., :3])
loss_R = huber_loss(cur_pred_pose_enc[..., 3:7], gt_pose_encoding[..., 3:7])
loss_fl = huber_loss(cur_pred_pose_enc[..., 7:], gt_pose_encoding[..., 7:])
else:
raise ValueError(f"Unknown loss type: {loss_type}")
loss_T = torch.nan_to_num(loss_T, nan=0.0, posinf=0.0, neginf=0.0)
loss_R = torch.nan_to_num(loss_R, nan=0.0, posinf=0.0, neginf=0.0)
loss_fl = torch.nan_to_num(loss_fl, nan=0.0, posinf=0.0, neginf=0.0)
loss_T = torch.clamp(loss_T, min=-100, max=100)
loss_R = torch.clamp(loss_R, min=-100, max=100)
loss_fl = torch.clamp(loss_fl, min=-100, max=100)
loss_T = loss_T.mean()
loss_R = loss_R.mean()
loss_fl = loss_fl.mean()
return loss_T, loss_R, loss_fl
def forward(self, distill_infos, pred_pose_enc_list, prediction, batch):
loss_pose = 0.0
if pred_pose_enc_list is not None:
num_predictions = len(pred_pose_enc_list)
pesudo_gt_pose_enc = distill_infos['pred_pose_enc_list']
for i in range(num_predictions):
i_weight = self.gamma ** (num_predictions - i - 1)
cur_pred_pose_enc = pred_pose_enc_list[i]
cur_pesudo_gt_pose_enc = pesudo_gt_pose_enc[i]
loss_pose += i_weight * huber_loss(cur_pred_pose_enc, cur_pesudo_gt_pose_enc).mean()
loss_pose = loss_pose / num_predictions
loss_pose = torch.nan_to_num(loss_pose, nan=0.0, posinf=0.0, neginf=0.0)
pred_depth = prediction.depth.flatten(0, 1)
pesudo_gt_depth = distill_infos['depth_map'].flatten(0, 1).squeeze(-1)
conf_mask = distill_infos['conf_mask'].flatten(0, 1)
if batch['context']['valid_mask'].sum() > 0:
conf_mask = batch['context']['valid_mask'].flatten(0, 1)
loss_depth = F.mse_loss(pred_depth[conf_mask], pesudo_gt_depth[conf_mask], reduction='none').mean()
render_normal = get_normal_map(pred_depth, batch["context"]["intrinsics"].flatten(0, 1))
pred_normal = get_normal_map(pesudo_gt_depth, batch["context"]["intrinsics"].flatten(0, 1))
alpha1_loss = (1 - (render_normal[conf_mask] * pred_normal[conf_mask]).sum(-1)).mean()
alpha2_loss = F.l1_loss(render_normal[conf_mask], pred_normal[conf_mask], reduction='mean')
loss_normal = (alpha1_loss + alpha2_loss) / 2
loss_distill = loss_pose * self.weight_pose + loss_depth * self.weight_depth + loss_normal * self.weight_normal
loss_distill = torch.nan_to_num(loss_distill, nan=0.0, posinf=0.0, neginf=0.0)
loss_dict = {
"loss_distill": loss_distill,
"loss_pose": loss_pose * self.weight_pose,
"loss_depth": loss_depth * self.weight_depth,
"loss_normal": loss_normal * self.weight_normal
}
return loss_dict