Spaces:
Running
Running
File size: 12,159 Bytes
a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 |
"""
Implements the full pipeline from raw images to line matches.
"""
import time
import cv2
import numpy as np
import torch
import torch.nn.functional as F
from torch.nn.functional import softmax
from .model_util import get_model
from .loss import get_loss_and_weights
from .metrics import super_nms
from .line_detection import LineSegmentDetectionModule
from .line_matching import WunschLineMatcher
from ..train import convert_junc_predictions
from ..misc.train_utils import adapt_checkpoint
from .line_detector import line_map_to_segments
class LineMatcher(object):
"""Full line matcher including line detection and matching
with the Needleman-Wunsch algorithm."""
def __init__(
self,
model_cfg,
ckpt_path,
device,
line_detector_cfg,
line_matcher_cfg,
multiscale=False,
scales=[1.0, 2.0],
):
# Get loss weights if dynamic weighting
_, loss_weights = get_loss_and_weights(model_cfg, device)
self.device = device
# Initialize the cnn backbone
self.model = get_model(model_cfg, loss_weights)
checkpoint = torch.load(ckpt_path, map_location=self.device)
checkpoint = adapt_checkpoint(checkpoint["model_state_dict"])
self.model.load_state_dict(checkpoint)
self.model = self.model.to(self.device)
self.model = self.model.eval()
self.grid_size = model_cfg["grid_size"]
self.junc_detect_thresh = model_cfg["detection_thresh"]
self.max_num_junctions = model_cfg.get("max_num_junctions", 300)
# Initialize the line detector
self.line_detector = LineSegmentDetectionModule(**line_detector_cfg)
self.multiscale = multiscale
self.scales = scales
# Initialize the line matcher
self.line_matcher = WunschLineMatcher(**line_matcher_cfg)
# Print some debug messages
for key, val in line_detector_cfg.items():
print(f"[Debug] {key}: {val}")
# print("[Debug] detect_thresh: %f" % (line_detector_cfg["detect_thresh"]))
# print("[Debug] num_samples: %d" % (line_detector_cfg["num_samples"]))
# Perform line detection and descriptor inference on a single image
def line_detection(
self, input_image, valid_mask=None, desc_only=False, profile=False
):
# Restrict input_image to 4D torch tensor
if (not len(input_image.shape) == 4) or (
not isinstance(input_image, torch.Tensor)
):
raise ValueError("[Error] the input image should be a 4D torch tensor")
# Move the input to corresponding device
input_image = input_image.to(self.device)
# Forward of the CNN backbone
start_time = time.time()
with torch.no_grad():
net_outputs = self.model(input_image)
outputs = {"descriptor": net_outputs["descriptors"]}
if not desc_only:
junc_np = convert_junc_predictions(
net_outputs["junctions"],
self.grid_size,
self.junc_detect_thresh,
self.max_num_junctions,
)
if valid_mask is None:
junctions = np.where(junc_np["junc_pred_nms"].squeeze())
else:
junctions = np.where(junc_np["junc_pred_nms"].squeeze() * valid_mask)
junctions = np.concatenate(
[junctions[0][..., None], junctions[1][..., None]], axis=-1
)
if net_outputs["heatmap"].shape[1] == 2:
# Convert to single channel directly from here
heatmap = (
softmax(net_outputs["heatmap"], dim=1)[:, 1:, :, :]
.cpu()
.numpy()
.transpose(0, 2, 3, 1)
)
else:
heatmap = (
torch.sigmoid(net_outputs["heatmap"])
.cpu()
.numpy()
.transpose(0, 2, 3, 1)
)
heatmap = heatmap[0, :, :, 0]
# Run the line detector.
line_map, junctions, heatmap = self.line_detector.detect(
junctions, heatmap, device=self.device
)
if isinstance(line_map, torch.Tensor):
line_map = line_map.cpu().numpy()
if isinstance(junctions, torch.Tensor):
junctions = junctions.cpu().numpy()
outputs["heatmap"] = heatmap.cpu().numpy()
outputs["junctions"] = junctions
# If it's a line map with multiple detect_thresh and inlier_thresh
if len(line_map.shape) > 2:
num_detect_thresh = line_map.shape[0]
num_inlier_thresh = line_map.shape[1]
line_segments = []
for detect_idx in range(num_detect_thresh):
line_segments_inlier = []
for inlier_idx in range(num_inlier_thresh):
line_map_tmp = line_map[detect_idx, inlier_idx, :, :]
line_segments_tmp = line_map_to_segments(
junctions, line_map_tmp
)
line_segments_inlier.append(line_segments_tmp)
line_segments.append(line_segments_inlier)
else:
line_segments = line_map_to_segments(junctions, line_map)
outputs["line_segments"] = line_segments
end_time = time.time()
if profile:
outputs["time"] = end_time - start_time
return outputs
# Perform line detection and descriptor inference at multiple scales
def multiscale_line_detection(
self,
input_image,
valid_mask=None,
desc_only=False,
profile=False,
scales=[1.0, 2.0],
aggregation="mean",
):
# Restrict input_image to 4D torch tensor
if (not len(input_image.shape) == 4) or (
not isinstance(input_image, torch.Tensor)
):
raise ValueError("[Error] the input image should be a 4D torch tensor")
# Move the input to corresponding device
input_image = input_image.to(self.device)
img_size = input_image.shape[2:4]
desc_size = tuple(np.array(img_size) // 4)
# Run the inference at multiple image scales
start_time = time.time()
junctions, heatmaps, descriptors = [], [], []
for s in scales:
# Resize the image
resized_img = F.interpolate(input_image, scale_factor=s, mode="bilinear")
# Forward of the CNN backbone
with torch.no_grad():
net_outputs = self.model(resized_img)
descriptors.append(
F.interpolate(
net_outputs["descriptors"], size=desc_size, mode="bilinear"
)
)
if not desc_only:
junc_prob = convert_junc_predictions(
net_outputs["junctions"], self.grid_size
)["junc_pred"]
junctions.append(
cv2.resize(
junc_prob.squeeze(),
(img_size[1], img_size[0]),
interpolation=cv2.INTER_LINEAR,
)
)
if net_outputs["heatmap"].shape[1] == 2:
# Convert to single channel directly from here
heatmap = softmax(net_outputs["heatmap"], dim=1)[:, 1:, :, :]
else:
heatmap = torch.sigmoid(net_outputs["heatmap"])
heatmaps.append(F.interpolate(heatmap, size=img_size, mode="bilinear"))
# Aggregate the results
if aggregation == "mean":
# Aggregation through the mean activation
descriptors = torch.stack(descriptors, dim=0).mean(0)
else:
# Aggregation through the max activation
descriptors = torch.stack(descriptors, dim=0).max(0)[0]
outputs = {"descriptor": descriptors}
if not desc_only:
if aggregation == "mean":
junctions = np.stack(junctions, axis=0).mean(0)[None]
heatmap = torch.stack(heatmaps, dim=0).mean(0)[0, 0, :, :]
heatmap = heatmap.cpu().numpy()
else:
junctions = np.stack(junctions, axis=0).max(0)[None]
heatmap = torch.stack(heatmaps, dim=0).max(0)[0][0, 0, :, :]
heatmap = heatmap.cpu().numpy()
# Extract junctions
junc_pred_nms = super_nms(
junctions[..., None],
self.grid_size,
self.junc_detect_thresh,
self.max_num_junctions,
)
if valid_mask is None:
junctions = np.where(junc_pred_nms.squeeze())
else:
junctions = np.where(junc_pred_nms.squeeze() * valid_mask)
junctions = np.concatenate(
[junctions[0][..., None], junctions[1][..., None]], axis=-1
)
# Run the line detector.
line_map, junctions, heatmap = self.line_detector.detect(
junctions, heatmap, device=self.device
)
if isinstance(line_map, torch.Tensor):
line_map = line_map.cpu().numpy()
if isinstance(junctions, torch.Tensor):
junctions = junctions.cpu().numpy()
outputs["heatmap"] = heatmap.cpu().numpy()
outputs["junctions"] = junctions
# If it's a line map with multiple detect_thresh and inlier_thresh
if len(line_map.shape) > 2:
num_detect_thresh = line_map.shape[0]
num_inlier_thresh = line_map.shape[1]
line_segments = []
for detect_idx in range(num_detect_thresh):
line_segments_inlier = []
for inlier_idx in range(num_inlier_thresh):
line_map_tmp = line_map[detect_idx, inlier_idx, :, :]
line_segments_tmp = line_map_to_segments(
junctions, line_map_tmp
)
line_segments_inlier.append(line_segments_tmp)
line_segments.append(line_segments_inlier)
else:
line_segments = line_map_to_segments(junctions, line_map)
outputs["line_segments"] = line_segments
end_time = time.time()
if profile:
outputs["time"] = end_time - start_time
return outputs
def __call__(self, images, valid_masks=[None, None], profile=False):
# Line detection and descriptor inference on both images
if self.multiscale:
forward_outputs = [
self.multiscale_line_detection(
images[0], valid_masks[0], profile=profile, scales=self.scales
),
self.multiscale_line_detection(
images[1], valid_masks[1], profile=profile, scales=self.scales
),
]
else:
forward_outputs = [
self.line_detection(images[0], valid_masks[0], profile=profile),
self.line_detection(images[1], valid_masks[1], profile=profile),
]
line_seg1 = forward_outputs[0]["line_segments"]
line_seg2 = forward_outputs[1]["line_segments"]
desc1 = forward_outputs[0]["descriptor"]
desc2 = forward_outputs[1]["descriptor"]
# Match the lines in both images
start_time = time.time()
matches = self.line_matcher.forward(line_seg1, line_seg2, desc1, desc2)
end_time = time.time()
outputs = {"line_segments": [line_seg1, line_seg2], "matches": matches}
if profile:
outputs["line_detection_time"] = (
forward_outputs[0]["time"] + forward_outputs[1]["time"]
)
outputs["line_matching_time"] = end_time - start_time
return outputs
|