colmap_to_mega_nerf.py

# All of the model reading methods are taken from https://github.com/cvg/Hierarchical-Localization, original
# license listed below:
#
# Copyright (c) 2018, ETH Zurich and UNC Chapel Hill.
# All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
#     * Redistributions of source code must retain the above copyright
#       notice, this list of conditions and the following disclaimer.
#
#     * Redistributions in binary form must reproduce the above copyright
#       notice, this list of conditions and the following disclaimer in the
#       documentation and/or other materials provided with the distribution.
#
#     * Neither the name of ETH Zurich and UNC Chapel Hill nor the names of
#       its contributors may be used to endorse or promote products derived
#       from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDERS OR CONTRIBUTORS BE
# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
# SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
# INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
# CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
# ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
# POSSIBILITY OF SUCH DAMAGE.
#
# Author: Johannes L. Schoenberger (jsch-at-demuc-dot-de)

import argparse
import collections
import os
import struct
from argparse import Namespace
from pathlib import Path

import cv2
import numpy as np
import torch
from tqdm import tqdm

import zipfile
from zipfile import ZipFile

RDF_TO_DRB = torch.FloatTensor([[0, 1, 0],
                                [1, 0, 0],
                                [0, 0, -1]])

CameraModel = collections.namedtuple(
    "CameraModel", ["model_id", "model_name", "num_params"])
Camera = collections.namedtuple(
    "Camera", ["id", "model", "width", "height", "params"])
BaseImage = collections.namedtuple(
    "Image", ["id", "qvec", "tvec", "camera_id", "name", "xys", "point3D_ids"])
Point3D = collections.namedtuple(
    "Point3D", ["id", "xyz", "rgb", "error", "image_ids", "point2D_idxs"])

CAMERA_MODELS = {
    CameraModel(model_id=0, model_name="SIMPLE_PINHOLE", num_params=3),
    CameraModel(model_id=1, model_name="PINHOLE", num_params=4),
    CameraModel(model_id=2, model_name="SIMPLE_RADIAL", num_params=4),
    CameraModel(model_id=3, model_name="RADIAL", num_params=5),
    CameraModel(model_id=4, model_name="OPENCV", num_params=8),
    CameraModel(model_id=5, model_name="OPENCV_FISHEYE", num_params=8),
    CameraModel(model_id=6, model_name="FULL_OPENCV", num_params=12),
    CameraModel(model_id=7, model_name="FOV", num_params=5),
    CameraModel(model_id=8, model_name="SIMPLE_RADIAL_FISHEYE", num_params=4),
    CameraModel(model_id=9, model_name="RADIAL_FISHEYE", num_params=5),
    CameraModel(model_id=10, model_name="THIN_PRISM_FISHEYE", num_params=12)
}
CAMERA_MODEL_IDS = dict([(camera_model.model_id, camera_model)
                         for camera_model in CAMERA_MODELS])


def qvec2rotmat(qvec):
    return np.array([
        [1 - 2 * qvec[2] ** 2 - 2 * qvec[3] ** 2,
         2 * qvec[1] * qvec[2] - 2 * qvec[0] * qvec[3],
         2 * qvec[3] * qvec[1] + 2 * qvec[0] * qvec[2]],
        [2 * qvec[1] * qvec[2] + 2 * qvec[0] * qvec[3],
         1 - 2 * qvec[1] ** 2 - 2 * qvec[3] ** 2,
         2 * qvec[2] * qvec[3] - 2 * qvec[0] * qvec[1]],
        [2 * qvec[3] * qvec[1] - 2 * qvec[0] * qvec[2],
         2 * qvec[2] * qvec[3] + 2 * qvec[0] * qvec[1],
         1 - 2 * qvec[1] ** 2 - 2 * qvec[2] ** 2]])


class Image(BaseImage):
    def qvec2rotmat(self):
        return qvec2rotmat(self.qvec)


def read_next_bytes(fid, num_bytes, format_char_sequence, endian_character="<"):
    """Read and unpack the next bytes from a binary file.
    :param fid:
    :param num_bytes: Sum of combination of {2, 4, 8}, e.g. 2, 6, 16, 30, etc.
    :param format_char_sequence: List of {c, e, f, d, h, H, i, I, l, L, q, Q}.
    :param endian_character: Any of {@, =, <, >, !}
    :return: Tuple of read and unpacked values.
    """
    data = fid.read(num_bytes)
    return struct.unpack(endian_character + format_char_sequence, data)


def read_points3D_binary(path_to_model_file):
    """
    see: src/base/reconstruction.cc
        void Reconstruction::ReadPoints3DBinary(const std::string& path)
        void Reconstruction::WritePoints3DBinary(const std::string& path)
    """
    points3D = {}
    with open(path_to_model_file, "rb") as fid:
        num_points = read_next_bytes(fid, 8, "Q")[0]
        for _ in range(num_points):
            binary_point_line_properties = read_next_bytes(
                fid, num_bytes=43, format_char_sequence="QdddBBBd")
            point3D_id = binary_point_line_properties[0]
            xyz = np.array(binary_point_line_properties[1:4])
            rgb = np.array(binary_point_line_properties[4:7])
            error = np.array(binary_point_line_properties[7])
            track_length = read_next_bytes(
                fid, num_bytes=8, format_char_sequence="Q")[0]
            track_elems = read_next_bytes(
                fid, num_bytes=8 * track_length,
                format_char_sequence="ii" * track_length)
            image_ids = np.array(tuple(map(int, track_elems[0::2])))
            point2D_idxs = np.array(tuple(map(int, track_elems[1::2])))
            points3D[point3D_id] = Point3D(
                id=point3D_id, xyz=xyz, rgb=rgb,
                error=error, image_ids=image_ids,
                point2D_idxs=point2D_idxs)
    return points3D


def read_images_binary(path_to_model_file):
    """
    see: src/base/reconstruction.cc
        void Reconstruction::ReadImagesBinary(const std::string& path)
        void Reconstruction::WriteImagesBinary(const std::string& path)
    """
    images = {}
    with open(path_to_model_file, "rb") as fid:
        num_reg_images = read_next_bytes(fid, 8, "Q")[0]
        for _ in range(num_reg_images):
            binary_image_properties = read_next_bytes(
                fid, num_bytes=64, format_char_sequence="idddddddi")
            image_id = binary_image_properties[0]
            qvec = np.array(binary_image_properties[1:5])
            tvec = np.array(binary_image_properties[5:8])
            camera_id = binary_image_properties[8]
            image_name = ""
            current_char = read_next_bytes(fid, 1, "c")[0]
            while current_char != b"\x00":  # look for the ASCII 0 entry
                image_name += current_char.decode("utf-8")
                current_char = read_next_bytes(fid, 1, "c")[0]
            num_points2D = read_next_bytes(fid, num_bytes=8,
                                           format_char_sequence="Q")[0]
            x_y_id_s = read_next_bytes(fid, num_bytes=24 * num_points2D,
                                       format_char_sequence="ddq" * num_points2D)
            xys = np.column_stack([tuple(map(float, x_y_id_s[0::3])),
                                   tuple(map(float, x_y_id_s[1::3]))])
            point3D_ids = np.array(tuple(map(int, x_y_id_s[2::3])))
            images[image_id] = Image(
                id=image_id, qvec=qvec, tvec=tvec,
                camera_id=camera_id, name=image_name,
                xys=xys, point3D_ids=point3D_ids)
    return images


def read_cameras_binary(path_to_model_file):
    """
    see: src/base/reconstruction.cc
        void Reconstruction::WriteCamerasBinary(const std::string& path)
        void Reconstruction::ReadCamerasBinary(const std::string& path)
    """
    cameras = {}
    with open(path_to_model_file, "rb") as fid:
        num_cameras = read_next_bytes(fid, 8, "Q")[0]
        for _ in range(num_cameras):
            camera_properties = read_next_bytes(
                fid, num_bytes=24, format_char_sequence="iiQQ")
            camera_id = camera_properties[0]
            model_id = camera_properties[1]
            model_name = CAMERA_MODEL_IDS[camera_properties[1]].model_name
            width = camera_properties[2]
            height = camera_properties[3]
            num_params = CAMERA_MODEL_IDS[model_id].num_params
            params = read_next_bytes(fid, num_bytes=8 * num_params,
                                     format_char_sequence="d" * num_params)
            cameras[camera_id] = Camera(id=camera_id,
                                        model=model_name,
                                        width=width,
                                        height=height,
                                        params=np.array(params))
        assert len(cameras) == num_cameras
    return cameras


def read_points3D_text(path):
    """
    see: src/base/reconstruction.cc
        void Reconstruction::ReadPoints3DText(const std::string& path)
        void Reconstruction::WritePoints3DText(const std::string& path)
    """
    points3D = {}
    with open(path, "r") as fid:
        while True:
            line = fid.readline()
            if not line:
                break
            line = line.strip()
            if len(line) > 0 and line[0] != "#":
                elems = line.split()
                point3D_id = int(elems[0])
                xyz = np.array(tuple(map(float, elems[1:4])))
                rgb = np.array(tuple(map(int, elems[4:7])))
                error = float(elems[7])
                image_ids = np.array(tuple(map(int, elems[8::2])))
                point2D_idxs = np.array(tuple(map(int, elems[9::2])))
                points3D[point3D_id] = Point3D(id=point3D_id, xyz=xyz, rgb=rgb,
                                               error=error, image_ids=image_ids,
                                               point2D_idxs=point2D_idxs)
    return points3D


def read_images_text(path):
    """
    see: src/base/reconstruction.cc
        void Reconstruction::ReadImagesText(const std::string& path)
        void Reconstruction::WriteImagesText(const std::string& path)
    """
    images = {}
    with open(path, "r") as fid:
        while True:
            line = fid.readline()
            if not line:
                break
            line = line.strip()
            if len(line) > 0 and line[0] != "#":
                elems = line.split()
                image_id = int(elems[0])
                qvec = np.array(tuple(map(float, elems[1:5])))
                tvec = np.array(tuple(map(float, elems[5:8])))
                camera_id = int(elems[8])
                image_name = elems[9]
                elems = fid.readline().split()
                xys = np.column_stack([tuple(map(float, elems[0::3])),
                                       tuple(map(float, elems[1::3]))])
                point3D_ids = np.array(tuple(map(int, elems[2::3])))
                images[image_id] = Image(
                    id=image_id, qvec=qvec, tvec=tvec,
                    camera_id=camera_id, name=image_name,
                    xys=xys, point3D_ids=point3D_ids)
    return images


def read_cameras_text(path):
    """
    see: src/base/reconstruction.cc
        void Reconstruction::WriteCamerasText(const std::string& path)
        void Reconstruction::ReadCamerasText(const std::string& path)
    """
    cameras = {}
    with open(path, "r") as fid:
        while True:
            line = fid.readline()
            if not line:
                break
            line = line.strip()
            if len(line) > 0 and line[0] != "#":
                elems = line.split()
                camera_id = int(elems[0])
                model = elems[1]
                width = int(elems[2])
                height = int(elems[3])
                params = np.array(tuple(map(float, elems[4:])))
                cameras[camera_id] = Camera(id=camera_id, model=model,
                                            width=width, height=height,
                                            params=params)
    return cameras


def detect_model_format(path, ext):
    if os.path.isfile(os.path.join(path, "cameras" + ext)) and \
            os.path.isfile(os.path.join(path, "images" + ext)) and \
            os.path.isfile(os.path.join(path, "points3D" + ext)):
        return True

    return False


def read_model(path, ext=""):
    # try to detect the extension automatically
    if ext == "":
        if detect_model_format(path, ".bin"):
            ext = ".bin"
        elif detect_model_format(path, ".txt"):
            ext = ".txt"
        else:
            try:
                cameras, images, points3D = read_model(os.path.join(path, "model/"))
                return cameras, images, points3D
            except FileNotFoundError:
                raise FileNotFoundError(
                    f"Could not find binary or text COLMAP model at {path}")

    if ext == ".txt":
        cameras = read_cameras_text(os.path.join(path, "cameras" + ext))
        images = read_images_text(os.path.join(path, "images" + ext))
        points3D = read_points3D_text(os.path.join(path, "points3D") + ext)
    else:
        cameras = read_cameras_binary(os.path.join(path, "cameras" + ext))
        images = read_images_binary(os.path.join(path, "images" + ext))
        points3D = read_points3D_binary(os.path.join(path, "points3D") + ext)
    return cameras, images, points3D


def _get_opts() -> Namespace:
    parser = argparse.ArgumentParser()

    parser.add_argument('--model_path', type=str, required=True, help='Path to PixSFM/COLMAP model')
    parser.add_argument('--images_path', type=str, required=True, help='Path to images')
    parser.add_argument('--output_path', type=str, required=True, help='Path to write converted dataset to')
    parser.add_argument('--scale', type=float, required=True,
                        help='Scale all poses by this factor. You generally want this to be between [-1, 1]')
    parser.add_argument('--num_val', type=int, default=20, help='Number of images to hold out in validation set')

    return parser.parse_args()


def main(hparams: Namespace) -> None:
    cameras, images, _ = read_model(hparams.model_path)

    c2ws = {}
    for image in images.values():
        w2c = torch.eye(4)
        w2c[:3, :3] = torch.FloatTensor(qvec2rotmat(image.qvec))
        w2c[:3, 3] = torch.FloatTensor(image.tvec)
        c2w = torch.inverse(w2c)

        c2w = torch.hstack((
            RDF_TO_DRB @ c2w[:3, :3] @ torch.inverse(RDF_TO_DRB),
            RDF_TO_DRB @ c2w[:3, 3:]
        ))

        c2ws[image.id] = c2w

    positions = torch.cat([c2w[:3, 3].unsqueeze(0) for c2w in c2ws.values()])
    print('{} images'.format(positions.shape[0]))
    max_values = positions.max(0)[0]
    min_values = positions.min(0)[0]
    origin = ((max_values + min_values) * 0.5)
    dist = (positions - origin).norm(dim=-1)
    diagonal = dist.max()

    print(origin, diagonal, max_values, min_values)
    coordinates = {
        'origin_drb': origin,
        'pose_scale_factor': hparams.scale
    }

    output_path = Path(hparams.output_path)
    output_path.mkdir(parents=True)
    (output_path / 'train' / 'metadata').mkdir(parents=True)
    (output_path / 'val' / 'metadata').mkdir(parents=True)

    (output_path / 'train' / 'rgbs').mkdir(parents=True)
    (output_path / 'val' / 'rgbs').mkdir(parents=True)

    images_path = Path(hparams.images_path)

    with (output_path / 'mappings.txt').open('w') as f:
        for i, image in enumerate(tqdm(sorted(images.values(), key=lambda x: x.name))):
            if i % int(positions.shape[0] / hparams.num_val) == 0:
                split_dir = output_path / 'val'
            else:
                split_dir = output_path / 'train'

            distorted = cv2.imread(str(images_path / image.name))

            camera = cameras[image.camera_id]

            # TODO: make camera model more flexible - should mainly involve changing the camera matrix accordingly
            assert camera.model == 'SIMPLE_RADIAL', camera.model

            camera_matrix = np.array([[camera.params[0], 0, camera.params[1]],
                                      [0, camera.params[0], camera.params[2]],
                                      [0, 0, 1]])

            distortion = np.array([camera.params[3], 0, 0, 0])
            undistorted = cv2.undistort(distorted, camera_matrix, distortion)
            cv2.imwrite(str(split_dir / 'rgbs' / '{0:06d}.jpg'.format(i)), undistorted)

            camera_in_drb = c2ws[image.id]
            camera_in_drb[:, 3] = (camera_in_drb[:, 3] - origin) / hparams.scale

            assert np.logical_and(camera_in_drb >= -1, camera_in_drb <= 1).all()

            metadata_name = '{0:06d}.pt'.format(i)
            torch.save({
                'H': distorted.shape[0],
                'W': distorted.shape[1],
                'c2w': torch.cat(
                    [camera_in_drb[:, 1:2], -camera_in_drb[:, :1], camera_in_drb[:, 2:4]],
                    -1),
                'intrinsics': torch.FloatTensor(
                    [camera_matrix[0][0], camera_matrix[1][1], camera_matrix[0][2], camera_matrix[1][2]]),
                'distortion': torch.FloatTensor(distortion)
            }, split_dir / 'metadata' / metadata_name)

            f.write('{},{}\n'.format(image.name, metadata_name))

    torch.save(coordinates, output_path / 'coordinates.pt')


if __name__ == '__main__':
    main(_get_opts())