Spaces:
Sleeping
Sleeping
File size: 8,208 Bytes
b74998d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
"""
Inference wrapper for DUSt3R
"""
import warnings
import torch
from dust3r.cloud_opt import global_aligner, GlobalAlignerMode
from dust3r.image_pairs import make_pairs
from dust3r.inference import inference
from dust3r.model import AsymmetricCroCo3DStereo # noqa
from mapanything.models.external.vggt.utils.rotation import mat_to_quat
from mapanything.utils.geometry import (
convert_ray_dirs_depth_along_ray_pose_trans_quats_to_pointmap,
convert_z_depth_to_depth_along_ray,
depthmap_to_camera_frame,
get_rays_in_camera_frame,
)
inf = float("inf")
def load_model(model_path, device, verbose=True):
if verbose:
print("Loading model from", model_path)
ckpt = torch.load(model_path, map_location="cpu", weights_only=False)
args = ckpt["args"].model.replace("ManyAR_PatchEmbed", "PatchEmbedDust3R")
if "landscape_only" not in args:
args = args[:-1] + ", landscape_only=False)"
else:
args = args.replace(" ", "").replace(
"landscape_only=True", "landscape_only=False"
)
assert "landscape_only=False" in args
if verbose:
print(f"Instantiating: {args}")
try:
net = eval(args)
except NameError:
net = AsymmetricCroCo3DStereo(
enc_depth=24,
dec_depth=12,
enc_embed_dim=1024,
dec_embed_dim=768,
enc_num_heads=16,
dec_num_heads=12,
pos_embed="RoPE100",
patch_embed_cls="PatchEmbedDust3R",
img_size=(512, 512),
head_type="dpt",
output_mode="pts3d",
depth_mode=("exp", -inf, inf),
conf_mode=("exp", 1, inf),
landscape_only=False,
)
s = net.load_state_dict(ckpt["model"], strict=False)
if verbose:
print(s)
return net.to(device)
class DUSt3RBAWrapper(torch.nn.Module):
def __init__(
self,
name,
ckpt_path,
scene_graph="complete",
inference_batch_size=32,
global_optim_schedule="cosine",
global_optim_lr=0.01,
global_optim_niter=300,
**kwargs,
):
super().__init__()
self.name = name
self.ckpt_path = ckpt_path
self.scene_graph = scene_graph
self.inference_batch_size = inference_batch_size
self.global_optim_schedule = global_optim_schedule
self.global_optim_lr = global_optim_lr
self.global_optim_niter = global_optim_niter
# Init the model and load the checkpoint
self.model = load_model(self.ckpt_path, device="cpu")
# Init the global aligner mode
self.global_aligner_mode = GlobalAlignerMode.PointCloudOptimizer
def forward(self, views):
"""
Forward pass wrapper for DUSt3R using the global aligner.
Assumption:
- The batch size of input views is 1.
Args:
views (List[dict]): List of dictionaries containing the input views' images and instance information.
Each dictionary should contain the following keys, where B is the batch size and is 1:
"img" (tensor): Image tensor of shape (B, C, H, W).
"data_norm_type" (list): ["dust3r"]
Returns:
List[dict]: A list containing the final outputs for the input views.
"""
# Check the batch size of input views
batch_size_per_view, _, height, width = views[0]["img"].shape
device = views[0]["img"].device
num_views = len(views)
assert batch_size_per_view == 1, (
f"Batch size of input views should be 1, but got {batch_size_per_view}."
)
# Check the data norm type
data_norm_type = views[0]["data_norm_type"][0]
assert data_norm_type == "dust3r", (
"DUSt3R expects a normalized image with the DUSt3R normalization scheme applied"
)
# Convert the input views to the expected input format
images = []
for view in views:
images.append(
dict(
img=view["img"],
idx=len(images),
instance=str(len(images)),
)
)
# Make image pairs and run inference pair-wise
pairs = make_pairs(
images, scene_graph=self.scene_graph, prefilter=None, symmetrize=True
)
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=FutureWarning)
output = inference(
pairs,
self.model,
device,
batch_size=self.inference_batch_size,
verbose=False,
)
# Global optimization
with torch.enable_grad():
scene = global_aligner(
output, device=device, mode=self.global_aligner_mode, verbose=False
)
_ = scene.compute_global_alignment(
init="mst",
niter=self.global_optim_niter,
schedule=self.global_optim_schedule,
lr=self.global_optim_lr,
)
# Make sure scene is not None
if scene is None:
raise RuntimeError("Global optimization failed.")
# Get the predictions
intrinsics = scene.get_intrinsics()
c2w_poses = scene.get_im_poses()
depths = scene.get_depthmaps()
# Convert the output to the MapAnything format
with torch.autocast("cuda", enabled=False):
res = []
for view_idx in range(num_views):
# Get the current view predictions
curr_view_intrinsic = intrinsics[view_idx].unsqueeze(0)
curr_view_pose = c2w_poses[view_idx].unsqueeze(0)
curr_view_depth_z = depths[view_idx].unsqueeze(0)
# Convert the pose to quaternions and translation
curr_view_cam_translations = curr_view_pose[..., :3, 3]
curr_view_cam_quats = mat_to_quat(curr_view_pose[..., :3, :3])
# Get the camera frame pointmaps
curr_view_pts3d_cam, _ = depthmap_to_camera_frame(
curr_view_depth_z, curr_view_intrinsic
)
# Convert the z depth to depth along ray
curr_view_depth_along_ray = convert_z_depth_to_depth_along_ray(
curr_view_depth_z, curr_view_intrinsic
)
curr_view_depth_along_ray = curr_view_depth_along_ray.unsqueeze(-1)
# Get the ray directions on the unit sphere in the camera frame
_, curr_view_ray_dirs = get_rays_in_camera_frame(
curr_view_intrinsic, height, width, normalize_to_unit_sphere=True
)
# Get the pointmaps
curr_view_pts3d = (
convert_ray_dirs_depth_along_ray_pose_trans_quats_to_pointmap(
curr_view_ray_dirs,
curr_view_depth_along_ray,
curr_view_cam_translations,
curr_view_cam_quats,
)
)
# Append the outputs to the result list
res.append(
{
"pts3d": curr_view_pts3d,
"pts3d_cam": curr_view_pts3d_cam,
"ray_directions": curr_view_ray_dirs,
"depth_along_ray": curr_view_depth_along_ray,
"cam_trans": curr_view_cam_translations,
"cam_quats": curr_view_cam_quats,
}
)
return res
|