Skip to content

Commit

Permalink
updated lightglue at the latest version to fix no keypoints crash
Browse files Browse the repository at this point in the history 
cvg/LightGlue#92
  • Loading branch information
@franioli
franioli committed Jan 11, 2024
1 parent a539fe6 commit ca93e67
Show file tree
Hide file tree
Showing 8 changed files with 230 additions and 195 deletions.
6 changes: 3 additions & 3 deletions .vscode/launch.json
View file
Original file line number Diff line number Diff line change
Expand Up @@ -16,19 +16,19 @@
"console": "integratedTerminal",
"args": [
"--dir",
"/home/francesco/casalbagliano/B_extended",
"/home/francesco/casalbagliano/C",
"--config",
"superpoint+lightglue",
"--strategy",
"bruteforce",
"matching_lowres",
// "--db_path",
// "/home/francesco/casalbagliano/B_extended/results_superpoint+lightglue_bruteforce_quality_high/reconstruction",
"--quality",
"highest",
"--tiling",
"preselection",
// "--skip_reconstruction",
"--force",
// "--force",
"-V"
]
},
Expand Down
Empty file added config.yaml
View file
Empty file.
323 changes: 159 additions & 164 deletions main_casalbagliano.py
View file
Original file line number Diff line number Diff line change
@@ -1,31 +1,78 @@
import sys
from importlib import import_module
from pathlib import Path
from venv import logger

from deep_image_matching import logger, timer
import pycolmap
from deep_image_matching import reconstruction, timer
from deep_image_matching.config import Config
from deep_image_matching.image_matching import ImageMatching
from deep_image_matching.io.h5_to_db import export_to_colmap
from deep_image_matching.parser import parse_cli

from scripts.metashape.metashape_from_dim import export_to_metashape

# Parse arguments from command line
args = parse_cli()

# Build configuration
config = Config(args)

# Updare some config parameters to not modify the main config file
config.general["tile_size"] = (3000, 2000)
config.general["gv_threshold"] = 4
config.general["gv_confidence"] = 0.99999
config.general["min_inliers_per_pair"] = 10
do_export_to_metashape = True

config.general["tile_size"] = (4500, 3000)
config.general["tile_overlap"] = 20
config.general["min_matches_per_tile"] = 10
config.general["tile_preselection_size"] = 1000
config.general["gv_threshold"] = 2
config.general["gv_confidence"] = 0.999999
config.general["min_inliers_per_pair"] = 30
config.general["min_inlier_ratio_per_pair"] = 0.3
config.general["try_match_full_images"] = True
config.extractor["max_keypoints"] = 1024
config.extractor["keypoint_threshold"] = 0.005
config.general["try_match_full_images"] = False

config.extractor["max_keypoints"] = 4000
config.extractor["keypoint_threshold"] = 0.05

config.matcher["n_layers"] = 9
config.matcher["depth_confidence"] = -1
config.matcher["width_confidence"] = -1
config.matcher["filter_threshold"] = 0.05
config.matcher["filter_threshold"] = 0.1
config.matcher["mp"] = True
config.matcher["flash"] = False

# Save configuration to file
config.save_config()

# Hard-coded parameters for Metashape
project_dir = config.general["output_dir"] / "metashape"
project_name = config.general["output_dir"].name + ".psx"
project_path = project_dir / project_name

rec_dir = config.general["output_dir"] / "reconstruction"
bundler_file_path = rec_dir / "bundler.out"
bundler_im_list = rec_dir / "bundler_list.txt"

marker_image_path = config.general["image_dir"].parent / "gcp_images_list.csv"
marker_world_path = config.general["image_dir"].parent / "gcp_list.csv"
column_format = "noxyz"

prm_to_optimize = {
"f": True,
"cx": True,
"cy": True,
"k1": True,
"k2": True,
"k3": True,
"k4": False,
"p1": True,
"p2": True,
"b1": False,
"b2": False,
"tiepoint_covariance": True,
}


existing_features_path = "/home/francesco/casalbagliano/C/results_superpoint+lightglue_matching_lowres_quality_highest/features.h5"
# existing_features_path = None

# For simplicity, save some of the configuration parameters in variables.
imgs_dir = config.general["image_dir"]
Expand All @@ -52,27 +99,42 @@
pair_path = img_matching.generate_pairs(num_matched=20)
timer.update("generate_pairs")

# Try to rotate images so they will be all "upright", useful for deep-learning approaches that usually are not rotation invariant
if config.general["upright"]:
img_matching.rotate_upright_images()
timer.update("rotate_upright_images")
if existing_features_path:
if not Path(existing_features_path).exists():
raise FileNotFoundError(
f"Features file {existing_features_path} does not exist"
)
feature_path = Path(existing_features_path)
logger.info(f"Using existing features from {feature_path}")
else:
# Check if features have already been extracted and remove it
if (output_dir / "features.h5").exists():
(output_dir / "features.h5").unlink()

# Try to rotate images so they will be all "upright", useful for deep-learning approaches that usually are not rotation invariant
if config.general["upright"]:
img_matching.rotate_upright_images()
timer.update("rotate_upright_images")

# Extract features
feature_path = img_matching.extract_features()
timer.update("extract_features")

# If features have been extracted on "upright" images, this function bring features back to their original image orientation
if config.general["upright"]:
img_matching.rotate_back_features(feature_path)
timer.update("rotate_back_features")

# Extract features
feature_path = img_matching.extract_features()
timer.update("extract_features")

# Matching
if (output_dir / "matches.h5").exists():
(output_dir / "matches.h5").unlink()
match_path = img_matching.match_pairs(
feature_path=feature_path,
try_full_image=config.general["try_match_full_images"],
)
timer.update("matching")

# If features have been extracted on "upright" images, this function bring features back to their original image orientation
if config.general["upright"]:
img_matching.rotate_back_features(feature_path)
timer.update("rotate_back_features")

# Export in colmap format
database_path = output_dir / "database.db"
export_to_colmap(
Expand All @@ -87,151 +149,84 @@

# If --skip_reconstruction is not specified, run reconstruction
if not config.general["skip_reconstruction"]:
use_pycolmap = True
try:
pycolmap = import_module("pycolmap")
except ImportError:
logger.error("Pycomlap is not available.")
use_pycolmap = False

if use_pycolmap:
# import reconstruction module
reconstruction = import_module("src.deep_image_matching.reconstruction")

# Define database path and camera mode
database = output_dir / "database_pycolmap.db"

# Define how pycolmap create the cameras. Possible CameraMode are:
# CameraMode.AUTO: infer the camera model based on the image exif
# CameraMode.PER_FOLDER: create a camera for each folder in the image directory
# CameraMode.PER_IMAGE: create a camera for each image in the image directory
# CameraMode.SINGLE: create a single camera for all images
camera_mode = pycolmap.CameraMode.AUTO

# Optional - You can manually define the cameras parameters (refer to https://github.com/colmap/colmap/blob/main/src/colmap/sensor/models.h).
# Note, that the cameras are first detected with the CameraMode specified and then overwitten with the custom model. Therefore, you MUST provide the SAME NUMBER of cameras and with the SAME ORDER in which the cameras appear in the COLMAP database.
# To see the camera number and order, you can run the reconstruction a first time with the AUTO camera mode (and without manually define the cameras) and see the list of cameras in the database with
# print(list(model.cameras.values()))
# or opening the database with the COLMAP gui.
#
# OPENCV camera models and number of parameters to be used
# SIMPLE_PINHOLE: f, cx, cy
# PINHOLE: fx, fy, cx, cy
# SIMPLE_RADIAL: f, cx, cy, k
# RADIAL: f, cx, cy, k1, k2
# OPENCV: fx, fy, cx, cy, k1, k2, p1, p2
# OPENCV_FISHEYE: fx, fy, cx, cy, k1, k2, k3, k4
# FULL_OPENCV: fx, fy, cx, cy, k1, k2, p1, p2, k3, k4, k5, k6
# FOV: fx, fy, cx, cy, omega
# SIMPLE_RADIAL_FISHEYE: f, cx, cy, k
# RADIAL_FISHEYE: f, cx, cy, k1, k2
# THIN_PRISM_FISHEYE: fx, fy, cx, cy, k1, k2, p1, p2, k3, k4, sx1, sy1
#
# cam0 = pycolmap.Camera(
# model="PINHOLE",
# width=1500,
# height=1000,
# params=[1500, 1500, 750, 500],
# )
# cam1 = pycolmap.Camera(
# model="SIMPLE_PINHOLE",
# width=6012,
# height=4008,
# params=[9.267, 3.053, 1.948],
# )
# cam2 = pycolmap.Camera(
# model="SIMPLE_PINHOLE",
# width=6012,
# height=4008,
# params=[6.621, 3.013, 1.943],
# )
# cameras = [cam0] # or cameras = [cam1, cam2]
cameras = [
pycolmap.Camera(
model="FULL_OPENCV",
width=5184,
height=3456,
params=[
4999.743761184519,
5000.6093606299955,
2624.437497520331,
1713.0874440334837,
-0.013354229348170842,
0.08536745182816073,
-5.6676843405109726e-05,
-0.0005887324415077316,
-0.15296510632186128,
0.0,
0.0,
0.0,
],
)
]
# cameras = None

# Optional - You can specify some reconstruction configuration
# reconst_opts = {
# "ba_refine_focal_length": True,
# "ba_refine_principal_point": False,
# "ba_refine_extra_params": True,
# }
reconst_opts = {}

# Run reconstruction
model = reconstruction.main(
database=database,
image_dir=imgs_dir,
feature_path=feature_path,
match_path=match_path,
pair_path=pair_path,
sfm_dir=output_dir,
camera_mode=camera_mode,
cameras=cameras,
skip_geometric_verification=False,
reconst_opts=reconst_opts,
verbose=config.general["verbose"],
# Define database path and camera mode
database = output_dir / "database_pycolmap.db"

# Define how pycolmap create the cameras. Possible CameraMode are:
# CameraMode.AUTO: infer the camera model based on the image exif
# CameraMode.PER_FOLDER: create a camera for each folder in the image directory
# CameraMode.PER_IMAGE: create a camera for each image in the image directory
# CameraMode.SINGLE: create a single camera for all images
camera_mode = pycolmap.CameraMode.AUTO

# Optional - You can manually define the cameras parameters (refer to https://github.com/colmap/colmap/blob/main/src/colmap/sensor/models.h).
# Note, that the cameras are first detected with the CameraMode specified and then overwitten with the custom model. Therefore, you MUST provide the SAME NUMBER of cameras and with the SAME ORDER in which the cameras appear in the COLMAP database.
# To see the camera number and order, you can run the reconstruction a first time with the AUTO camera mode (and without manually define the cameras) and see the list of cameras in the database with
# print(list(model.cameras.values()))
# or opening the database with the COLMAP gui.
#
# OPENCV camera models and number of parameters to be used
# SIMPLE_PINHOLE: f, cx, cy
# PINHOLE: fx, fy, cx, cy
# SIMPLE_RADIAL: f, cx, cy, k
# RADIAL: f, cx, cy, k1, k2
# OPENCV: fx, fy, cx, cy, k1, k2, p1, p2
# OPENCV_FISHEYE: fx, fy, cx, cy, k1, k2, k3, k4
# FULL_OPENCV: fx, fy, cx, cy, k1, k2, p1, p2, k3, k4, k5, k6
# FOV: fx, fy, cx, cy, omega
# SIMPLE_RADIAL_FISHEYE: f, cx, cy, k
# RADIAL_FISHEYE: f, cx, cy, k1, k2
# THIN_PRISM_FISHEYE: fx, fy, cx, cy, k1, k2, p1, p2, k3, k4, sx1, sy1

cameras = [
pycolmap.Camera(
model="FULL_OPENCV",
width=5184,
height=3456,
params=[
4999.743761184519,
5000.6093606299955,
2624.437497520331,
1713.0874440334837,
-0.013354229348170842,
0.08536745182816073,
-5.6676843405109726e-05,
-0.0005887324415077316,
-0.15296510632186128,
0.0,
0.0,
0.0,
],
)
]
# cameras = None

# Optional - You can specify some reconstruction configuration
reconst_opts = {
"ba_refine_focal_length": False,
"ba_refine_principal_point": False,
"ba_refine_extra_params": False,
}
# reconst_opts = {}

# Run reconstruction
model = reconstruction.main(
database=database,
image_dir=imgs_dir,
feature_path=feature_path,
match_path=match_path,
pair_path=pair_path,
sfm_dir=output_dir,
camera_mode=camera_mode,
cameras=cameras,
skip_geometric_verification=False,
reconst_opts=reconst_opts,
verbose=config.general["verbose"],
)

timer.update("pycolmap reconstruction")

else:
logger.warning("Reconstruction with COLMAP CLI is not implemented yet.")

do_export_to_metashape = True
timer.update("pycolmap reconstruction")

if model and do_export_to_metashape:
sys.path.append(Path(__file__))

from scripts.metashape.metashape_from_dim import export_to_metashape

# Hard-coded parameters
project_dir = config.general["output_dir"].parent / "metashape"
project_name = config.general["output_dir"].name + ".psx"
project_path = project_dir / project_name

rec_dir = config.general["output_dir"] / "reconstruction"
bundler_file_path = rec_dir / "bundler.out"
bundler_im_list = rec_dir / "bundler_list.txt"

marker_image_path = config.general["image_dir"].parent / "gcp_images_list.csv"
marker_world_path = config.general["image_dir"].parent / "gcp_list.csv"
column_format = "noxyz"

prm_to_optimize = {
"f": True,
"cx": True,
"cy": True,
"k1": True,
"k2": True,
"k3": True,
"k4": False,
"p1": True,
"p2": True,
"b1": False,
"b2": False,
"tiepoint_covariance": True,
}

export_to_metashape(
project_path=project_path,
images_dir=config.general["image_dir"],
Expand Down
0 scripts/metashape/export_gcps.py → scripts/metashape/export_gcps_.py
View file
File renamed without changes.
Loading

0 comments on commit ca93e67

Please sign in to comment.