cvat/utils/dataset_manifest/utils.py

# Copyright (C) 2021-2022 Intel Corporation
#
# SPDX-License-Identifier: MIT

import hashlib
import mimetypes
import os
import re
from enum import Enum
from random import shuffle

import cv2 as cv
from av import VideoFrame
from natsort import os_sorted


def rotate_image(image, angle):
    height, width = image.shape[:2]
    image_center = (width / 2, height / 2)
    matrix = cv.getRotationMatrix2D(image_center, angle, 1.0)
    abs_cos = abs(matrix[0, 0])
    abs_sin = abs(matrix[0, 1])
    bound_w = int(height * abs_sin + width * abs_cos)
    bound_h = int(height * abs_cos + width * abs_sin)
    matrix[0, 2] += bound_w / 2 - image_center[0]
    matrix[1, 2] += bound_h / 2 - image_center[1]
    matrix = cv.warpAffine(image, matrix, (bound_w, bound_h))
    return matrix


def md5_hash(frame):
    if isinstance(frame, VideoFrame):
        frame = frame.to_image()
    return hashlib.md5(frame.tobytes()).hexdigest()  # nosec


def _define_data_type(media):
    return mimetypes.guess_type(media)[0]


def is_video(media_file):
    data_type = _define_data_type(media_file)
    return data_type is not None and data_type.startswith("video")


def is_image(media_file):
    data_type = _define_data_type(media_file)
    return (
        data_type is not None
        and data_type.startswith("image")
        and not data_type.startswith("image/svg")
    )


def _list_and_join(root):
    files = os.listdir(root)
    for f in files:
        yield os.path.join(root, f)


def _prepare_context_list(files, base_dir):
    return sorted(os.path.relpath(x, base_dir) for x in filter(is_image, files))


# Expected 2D format is:
# data/
#   00001.png
#   related_images/
#     00001_png/
#       context_image_1.jpeg
#       context_image_2.png
def _detect_related_images_2D(image_paths, root_path):
    related_images = {}
    latest_dirname = ""
    related_images_exist = False

    for image_path in sorted(image_paths):
        rel_image_path = os.path.relpath(image_path, root_path)
        dirname = os.path.dirname(image_path)
        related_images_dirname = os.path.join(dirname, "related_images")
        related_images[rel_image_path] = []

        if latest_dirname == dirname and not related_images_exist:
            continue
        elif latest_dirname != dirname:
            # Update some data applicable for a subset of paths (within the current dirname)
            latest_dirname = dirname
            related_images_exist = os.path.isdir(related_images_dirname)

        if related_images_exist:
            related_images_dirname = os.path.join(
                related_images_dirname, "_".join(os.path.basename(image_path).rsplit(".", 1))
            )

            if os.path.isdir(related_images_dirname):
                related_images[rel_image_path] = _prepare_context_list(
                    _list_and_join(related_images_dirname), root_path
                )
    return related_images


# Possible 3D formats are:
# velodyne_points/
#     data/
#         image_01.bin
# IMAGE_00 # any number?
#     data/
#         image_01.png
#
# pointcloud/
#     00001.pcd
# related_images/
#     00001_pcd/
#         image_01.png # or other image
#
# Default formats
# Option 1
# data/
#     image.pcd
#     image.png
#
# Option 2
# data/
#    image_1/
#        image_1.pcd
#        context_1.png
#        context_2.jpg
def _detect_related_images_3D(image_paths, root_path):
    related_images = {}
    latest_dirname = ""
    dirname_files = []
    related_images_exist = False
    velodyne_context_images_dirs = []

    for image_path in sorted(image_paths):
        rel_image_path = os.path.relpath(image_path, root_path)
        name = os.path.splitext(os.path.basename(image_path))[0]
        dirname = os.path.dirname(image_path)
        related_images_dirname = os.path.normpath(os.path.join(dirname, "..", "related_images"))
        related_images[rel_image_path] = []

        if latest_dirname != dirname:
            # Update some data applicable for a subset of paths (within the current dirname)
            latest_dirname = dirname
            related_images_exist = os.path.isdir(related_images_dirname)
            dirname_files = list(_list_and_join(dirname))
            velodyne_context_images_dirs = [
                directory
                for directory in _list_and_join(os.path.normpath(os.path.join(dirname, "..", "..")))
                if os.path.isdir(os.path.join(directory, "data"))
                and re.search(r"image_\d.*", directory, re.IGNORECASE)
            ]

        filtered_dirname_files = list(filter(lambda x: x != image_path, dirname_files))
        if len(filtered_dirname_files) and os.path.basename(dirname) == name:
            # default format (option 2)
            related_images[rel_image_path].extend(
                _prepare_context_list(filtered_dirname_files, root_path)
            )
        else:
            filtered_dirname_files = list(
                filter(
                    lambda x: os.path.splitext(os.path.basename(x))[0] == name,
                    filtered_dirname_files,
                )
            )
            if len(filtered_dirname_files):
                # default format (option 1)
                related_images[rel_image_path].extend(
                    _prepare_context_list(filtered_dirname_files, root_path)
                )

        if related_images_exist:
            related_images_dirname = os.path.join(
                related_images_dirname, "_".join(os.path.basename(image_path).rsplit(".", 1))
            )
            if os.path.isdir(related_images_dirname):
                related_images[rel_image_path].extend(
                    _prepare_context_list(_list_and_join(related_images_dirname), root_path)
                )

        if dirname.endswith(os.path.join("velodyne_points", "data")):
            # velodynepoints format
            for context_images_dir in velodyne_context_images_dirs:
                context_files = _list_and_join(os.path.join(context_images_dir, "data"))
                context_files = list(
                    filter(
                        lambda x: os.path.splitext(os.path.basename(x))[0] == name, context_files
                    )
                )
                related_images[rel_image_path].extend(
                    _prepare_context_list(context_files, root_path)
                )

        related_images[rel_image_path].sort()
    return related_images


# This function is expected to be called only for images tasks
# image_path is expected to be a list of absolute path to images
# root_path is expected to be a string (dataset root)
def detect_related_images(image_paths, root_path):
    data_are_2d = False
    data_are_3d = False

    # First of all need to define data type we are working with
    for image_path in image_paths:
        # .bin files are expected to be converted to .pcd before this code
        if os.path.splitext(image_path)[1].lower() == ".pcd":
            data_are_3d = True
        else:
            data_are_2d = True
    assert not (data_are_3d and data_are_2d), "Combined data types 2D and 3D are not supported"

    if data_are_2d:
        return _detect_related_images_2D(image_paths, root_path)
    elif data_are_3d:
        return _detect_related_images_3D(image_paths, root_path)
    return {}


class SortingMethod(str, Enum):
    LEXICOGRAPHICAL = "lexicographical"
    NATURAL = "natural"
    PREDEFINED = "predefined"
    RANDOM = "random"

    @classmethod
    def choices(cls):
        return tuple((x.value, x.name) for x in cls)

    def __str__(self):
        return self.value


def sort(images, sorting_method=SortingMethod.LEXICOGRAPHICAL, func=None):
    if sorting_method == SortingMethod.LEXICOGRAPHICAL:
        return sorted(images, key=func)
    elif sorting_method == SortingMethod.NATURAL:
        return os_sorted(images, key=func)
    elif sorting_method == SortingMethod.PREDEFINED:
        return images
    elif sorting_method == SortingMethod.RANDOM:
        shuffle(images)
        return images
    else:
        raise NotImplementedError()