Spaces:
Running
on
Zero
Running
on
Zero
File size: 9,783 Bytes
a51c6d2 |
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 |
import os
os.environ['OPENCV_IO_ENABLE_OPENEXR'] = '1'
from pathlib import Path
from typing import *
import itertools
import json
import warnings
import cv2
import numpy as np
from numpy import ndarray
from tqdm import tqdm, trange
from scipy.sparse import csr_array, hstack, vstack
from scipy.ndimage import convolve
from scipy.sparse.linalg import lsmr
import utils3d
def get_panorama_cameras():
vertices, _ = utils3d.numpy.icosahedron()
intrinsics = utils3d.numpy.intrinsics_from_fov(fov_x=np.deg2rad(90), fov_y=np.deg2rad(90))
extrinsics = utils3d.numpy.extrinsics_look_at([0, 0, 0], vertices, [0, 0, 1]).astype(np.float32)
return extrinsics, [intrinsics] * len(vertices)
def spherical_uv_to_directions(uv: np.ndarray):
theta, phi = (1 - uv[..., 0]) * (2 * np.pi), uv[..., 1] * np.pi
directions = np.stack([np.sin(phi) * np.cos(theta), np.sin(phi) * np.sin(theta), np.cos(phi)], axis=-1)
return directions
def directions_to_spherical_uv(directions: np.ndarray):
directions = directions / np.linalg.norm(directions, axis=-1, keepdims=True)
u = 1 - np.arctan2(directions[..., 1], directions[..., 0]) / (2 * np.pi) % 1.0
v = np.arccos(directions[..., 2]) / np.pi
return np.stack([u, v], axis=-1)
def split_panorama_image(image: np.ndarray, extrinsics: np.ndarray, intrinsics: np.ndarray, resolution: int):
height, width = image.shape[:2]
uv = utils3d.numpy.image_uv(width=resolution, height=resolution)
splitted_images = []
for i in range(len(extrinsics)):
spherical_uv = directions_to_spherical_uv(utils3d.numpy.unproject_cv(uv, extrinsics=extrinsics[i], intrinsics=intrinsics[i]))
pixels = utils3d.numpy.uv_to_pixel(spherical_uv, width=width, height=height).astype(np.float32)
splitted_image = cv2.remap(image, pixels[..., 0], pixels[..., 1], interpolation=cv2.INTER_LINEAR)
splitted_images.append(splitted_image)
return splitted_images
def poisson_equation(width: int, height: int, wrap_x: bool = False, wrap_y: bool = False) -> Tuple[csr_array, ndarray]:
grid_index = np.arange(height * width).reshape(height, width)
grid_index = np.pad(grid_index, ((0, 0), (1, 1)), mode='wrap' if wrap_x else 'edge')
grid_index = np.pad(grid_index, ((1, 1), (0, 0)), mode='wrap' if wrap_y else 'edge')
data = np.array([[-4, 1, 1, 1, 1]], dtype=np.float32).repeat(height * width, axis=0).reshape(-1)
indices = np.stack([
grid_index[1:-1, 1:-1],
grid_index[:-2, 1:-1], # up
grid_index[2:, 1:-1], # down
grid_index[1:-1, :-2], # left
grid_index[1:-1, 2:] # right
], axis=-1).reshape(-1)
indptr = np.arange(0, height * width * 5 + 1, 5)
A = csr_array((data, indices, indptr), shape=(height * width, height * width))
return A
def grad_equation(width: int, height: int, wrap_x: bool = False, wrap_y: bool = False) -> Tuple[csr_array, np.ndarray]:
grid_index = np.arange(width * height).reshape(height, width)
if wrap_x:
grid_index = np.pad(grid_index, ((0, 0), (0, 1)), mode='wrap')
if wrap_y:
grid_index = np.pad(grid_index, ((0, 1), (0, 0)), mode='wrap')
data = np.concatenate([
np.concatenate([
np.ones((grid_index.shape[0], grid_index.shape[1] - 1), dtype=np.float32).reshape(-1, 1), # x[i,j]
-np.ones((grid_index.shape[0], grid_index.shape[1] - 1), dtype=np.float32).reshape(-1, 1), # x[i,j-1]
], axis=1).reshape(-1),
np.concatenate([
np.ones((grid_index.shape[0] - 1, grid_index.shape[1]), dtype=np.float32).reshape(-1, 1), # x[i,j]
-np.ones((grid_index.shape[0] - 1, grid_index.shape[1]), dtype=np.float32).reshape(-1, 1), # x[i-1,j]
], axis=1).reshape(-1),
])
indices = np.concatenate([
np.concatenate([
grid_index[:, :-1].reshape(-1, 1),
grid_index[:, 1:].reshape(-1, 1),
], axis=1).reshape(-1),
np.concatenate([
grid_index[:-1, :].reshape(-1, 1),
grid_index[1:, :].reshape(-1, 1),
], axis=1).reshape(-1),
])
indptr = np.arange(0, grid_index.shape[0] * (grid_index.shape[1] - 1) * 2 + (grid_index.shape[0] - 1) * grid_index.shape[1] * 2 + 1, 2)
A = csr_array((data, indices, indptr), shape=(grid_index.shape[0] * (grid_index.shape[1] - 1) + (grid_index.shape[0] - 1) * grid_index.shape[1], height * width))
return A
def merge_panorama_depth(width: int, height: int, distance_maps: List[np.ndarray], pred_masks: List[np.ndarray], extrinsics: List[np.ndarray], intrinsics: List[np.ndarray]):
if max(width, height) > 256:
panorama_depth_init, _ = merge_panorama_depth(width // 2, height // 2, distance_maps, pred_masks, extrinsics, intrinsics)
panorama_depth_init = cv2.resize(panorama_depth_init, (width, height), cv2.INTER_LINEAR)
else:
panorama_depth_init = None
uv = utils3d.numpy.image_uv(width=width, height=height)
spherical_directions = spherical_uv_to_directions(uv)
# Warp each view to the panorama
panorama_log_distance_grad_maps, panorama_grad_masks = [], []
panorama_log_distance_laplacian_maps, panorama_laplacian_masks = [], []
panorama_pred_masks = []
for i in range(len(distance_maps)):
projected_uv, projected_depth = utils3d.numpy.project_cv(spherical_directions, extrinsics=extrinsics[i], intrinsics=intrinsics[i])
projection_valid_mask = (projected_depth > 0) & (projected_uv > 0).all(axis=-1) & (projected_uv < 1).all(axis=-1)
projected_pixels = utils3d.numpy.uv_to_pixel(np.clip(projected_uv, 0, 1), width=distance_maps[i].shape[1], height=distance_maps[i].shape[0]).astype(np.float32)
log_splitted_distance = np.log(distance_maps[i])
panorama_log_distance_map = np.where(projection_valid_mask, cv2.remap(log_splitted_distance, projected_pixels[..., 0], projected_pixels[..., 1], cv2.INTER_LINEAR, borderMode=cv2.BORDER_REPLICATE), 0)
panorama_pred_mask = projection_valid_mask & (cv2.remap(pred_masks[i].astype(np.uint8), projected_pixels[..., 0], projected_pixels[..., 1], cv2.INTER_NEAREST, borderMode=cv2.BORDER_REPLICATE) > 0)
# calculate gradient map
padded = np.pad(panorama_log_distance_map, ((0, 0), (0, 1)), mode='wrap')
grad_x, grad_y = padded[:, :-1] - padded[:, 1:], padded[:-1, :] - padded[1:, :]
padded = np.pad(panorama_pred_mask, ((0, 0), (0, 1)), mode='wrap')
mask_x, mask_y = padded[:, :-1] & padded[:, 1:], padded[:-1, :] & padded[1:, :]
panorama_log_distance_grad_maps.append((grad_x, grad_y))
panorama_grad_masks.append((mask_x, mask_y))
# calculate laplacian map
padded = np.pad(panorama_log_distance_map, ((1, 1), (0, 0)), mode='edge')
padded = np.pad(padded, ((0, 0), (1, 1)), mode='wrap')
laplacian = convolve(padded, np.array([[0, 1, 0], [1, -4, 1], [0, 1, 0]], dtype=np.float32))[1:-1, 1:-1]
padded = np.pad(panorama_pred_mask, ((1, 1), (0, 0)), mode='edge')
padded = np.pad(padded, ((0, 0), (1, 1)), mode='wrap')
mask = convolve(padded.astype(np.uint8), np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.uint8))[1:-1, 1:-1] == 5
panorama_log_distance_laplacian_maps.append(laplacian)
panorama_laplacian_masks.append(mask)
panorama_pred_masks.append(panorama_pred_mask)
panorama_log_distance_grad_x = np.stack([grad_map[0] for grad_map in panorama_log_distance_grad_maps], axis=0)
panorama_log_distance_grad_y = np.stack([grad_map[1] for grad_map in panorama_log_distance_grad_maps], axis=0)
panorama_grad_mask_x = np.stack([mask_map[0] for mask_map in panorama_grad_masks], axis=0)
panorama_grad_mask_y = np.stack([mask_map[1] for mask_map in panorama_grad_masks], axis=0)
panorama_log_distance_grad_x = np.sum(panorama_log_distance_grad_x * panorama_grad_mask_x, axis=0) / np.sum(panorama_grad_mask_x, axis=0).clip(1e-3)
panorama_log_distance_grad_y = np.sum(panorama_log_distance_grad_y * panorama_grad_mask_y, axis=0) / np.sum(panorama_grad_mask_y, axis=0).clip(1e-3)
panorama_laplacian_maps = np.stack(panorama_log_distance_laplacian_maps, axis=0)
panorama_laplacian_masks = np.stack(panorama_laplacian_masks, axis=0)
panorama_laplacian_map = np.sum(panorama_laplacian_maps * panorama_laplacian_masks, axis=0) / np.sum(panorama_laplacian_masks, axis=0).clip(1e-3)
grad_x_mask = np.any(panorama_grad_mask_x, axis=0).reshape(-1)
grad_y_mask = np.any(panorama_grad_mask_y, axis=0).reshape(-1)
grad_mask = np.concatenate([grad_x_mask, grad_y_mask])
laplacian_mask = np.any(panorama_laplacian_masks, axis=0).reshape(-1)
# Solve overdetermined system
A = vstack([
grad_equation(width, height, wrap_x=True, wrap_y=False)[grad_mask],
poisson_equation(width, height, wrap_x=True, wrap_y=False)[laplacian_mask],
])
b = np.concatenate([
panorama_log_distance_grad_x.reshape(-1)[grad_x_mask],
panorama_log_distance_grad_y.reshape(-1)[grad_y_mask],
panorama_laplacian_map.reshape(-1)[laplacian_mask]
])
x, *_ = lsmr(
A, b,
atol=1e-5, btol=1e-5,
x0=np.log(panorama_depth_init).reshape(-1) if panorama_depth_init is not None else None,
show=False,
)
panorama_depth = np.exp(x).reshape(height, width).astype(np.float32)
panorama_mask = np.any(panorama_pred_masks, axis=0)
return panorama_depth, panorama_mask
|