@volter/blender-engine 0.1.147 → 0.1.149

package/browser/addons/vrm/common/human_bone_mapper/structure_based_mapping.py

@@ -0,0 +1,891 @@
+# SPDX-License-Identifier: MIT OR GPL-3.0-or-later
+
+import itertools
+import math
+from collections.abc import Iterator, Mapping
+from dataclasses import dataclass, field
+from functools import cached_property
+from typing import Final, Optional, Union
+
+from bpy.types import Armature, Bone, Object
+from mathutils import Matrix
+
+from ..deep import Json
+from ..logger import get_logger
+from ..vrm1.human_bone import (
+    HumanBoneName,
+    HumanBoneSpecification,
+    HumanBoneSpecifications,
+)
+
+_logger = get_logger(__name__)
+
+
+CENTRAL_BONE_NAMES = frozenset(
+    [
+        HumanBoneName.HEAD,
+        HumanBoneName.NECK,
+        HumanBoneName.UPPER_CHEST,
+        HumanBoneName.CHEST,
+        HumanBoneName.SPINE,
+        HumanBoneName.HIPS,
+    ]
+)
+
+HIPS_TO_SPINE_BONE_NAMES = frozenset(
+    [
+        HumanBoneName.HIPS,
+        HumanBoneName.SPINE,
+    ]
+)
+
+SPINE_TO_UPPER_CHEST_BONE_NAMES = frozenset(
+    [
+        HumanBoneName.SPINE,
+        HumanBoneName.CHEST,
+        HumanBoneName.UPPER_CHEST,
+    ]
+)
+
+SPINE_TO_NECK_NAMES = frozenset(
+    [
+        HumanBoneName.SPINE,
+        HumanBoneName.CHEST,
+        HumanBoneName.UPPER_CHEST,
+        HumanBoneName.NECK,
+    ]
+)
+
+LEFT_ARM_BONE_NAMES = frozenset(
+    [
+        HumanBoneName.LEFT_EYE,
+        HumanBoneName.LEFT_HAND,
+        HumanBoneName.LEFT_SHOULDER,
+        HumanBoneName.LEFT_UPPER_ARM,
+        HumanBoneName.LEFT_LOWER_ARM,
+    ]
+)
+
+LEFT_LEG_BONE_NAMES = frozenset(
+    [
+        HumanBoneName.LEFT_UPPER_LEG,
+        HumanBoneName.LEFT_LOWER_LEG,
+        HumanBoneName.LEFT_FOOT,
+    ]
+)
+
+RIGHT_ARM_BONE_NAMES = frozenset(
+    [
+        HumanBoneName.RIGHT_EYE,
+        HumanBoneName.RIGHT_HAND,
+        HumanBoneName.RIGHT_SHOULDER,
+        HumanBoneName.RIGHT_UPPER_ARM,
+        HumanBoneName.RIGHT_LOWER_ARM,
+    ]
+)
+
+RIGHT_LEG_BONE_NAMES = frozenset(
+    [
+        HumanBoneName.RIGHT_UPPER_LEG,
+        HumanBoneName.RIGHT_LOWER_LEG,
+        HumanBoneName.RIGHT_FOOT,
+    ]
+)
+
+
+@dataclass
+class Counter:
+    count: int = 0
+    cache_hit: int = 0
+
+
+@dataclass(frozen=True)
+class NormalizedBone:
+    name: str
+    x: float
+    y: float
+    z: float
+    children: tuple["NormalizedBone", ...]
+    recursive_len: int
+
+    @classmethod
+    def create(
+        cls,
+        armature_matrix: Matrix,
+        max_x: float,
+        max_y: float,
+        max_z: float,
+        bone: Bone,
+    ) -> "NormalizedBone":
+        x, y, z = (armature_matrix @ bone.matrix_local).translation
+        normalized_x = math.copysign(math.sqrt(abs(x) / max_x), x) if max_x > 0 else 0
+        normalized_y = math.copysign(math.sqrt(abs(y) / max_y), y) if max_y > 0 else 0
+        normalized_z = math.copysign(math.sqrt(abs(z) / max_z), z) if max_z > 0 else 0
+        children = tuple(
+            cls.create(armature_matrix, max_x, max_y, max_z, child)
+            for child in bone.children
+        )
+        return NormalizedBone(
+            name=bone.name,
+            x=normalized_x,
+            y=normalized_y,
+            z=normalized_z,
+            children=children,
+            recursive_len=1 + sum(child.recursive_len for child in children),
+        )
+
+    @classmethod
+    def from_json(cls, data: Json) -> "NormalizedBone":
+        if not isinstance(data, Mapping):
+            message = "NormalizedBone must be an object"
+            raise TypeError(message)
+        name = data.get("name")
+        if not isinstance(name, str):
+            message = "NormalizedBone.name must be a string"
+            raise TypeError(message)
+
+        x = data.get("x")
+        y = data.get("y")
+        z = data.get("z")
+        if not isinstance(x, (int, float)):
+            message = "NormalizedBone.x must be a number"
+            raise TypeError(message)
+        if not isinstance(y, (int, float)):
+            message = "NormalizedBone.y must be a number"
+            raise TypeError(message)
+        if not isinstance(z, (int, float)):
+            message = "NormalizedBone.z must be a number"
+            raise TypeError(message)
+
+        children_data = data.get("children")
+        if children_data is None:
+            children_data = []
+        if not isinstance(children_data, list):
+            message = "NormalizedBone.children must be a list"
+            raise TypeError(message)
+        children = tuple(cls.from_json(child) for child in children_data)
+
+        recursive_len = data.get("recursiveLen")
+        if recursive_len is None:
+            recursive_len_value = 1 + sum(child.recursive_len for child in children)
+        elif isinstance(recursive_len, int):
+            recursive_len_value = recursive_len
+        else:
+            message = "NormalizedBone.recursiveLen must be an int"
+            raise ValueError(message)
+
+        return NormalizedBone(
+            name=name,
+            x=float(x),
+            y=float(y),
+            z=float(z),
+            children=children,
+            recursive_len=recursive_len_value,
+        )
+
+    def to_json(self) -> Json:
+        return {
+            "name": self.name,
+            "x": self.x,
+            "y": self.y,
+            "z": self.z,
+            "children": [child.to_json() for child in self.children],
+            "recursiveLen": self.recursive_len,
+        }
+
+    def dump(self) -> str:
+        output = f'("{self.name}", ({self.x:.3f}, {self.y:.3f}, {self.z:.3f})): {{\n'
+        for child in self.children:
+            output += child.dump()
+        output += "},\n"
+        return output
+
+
+SearchBranch = Union["UnassignedSearchBranch", "AssignedSearchBranch"]
+
+
+def _parent_from_json(
+    data: Json,
+) -> Optional[tuple[NormalizedBone, HumanBoneSpecification]]:
+    if data is None:
+        return None
+    if not isinstance(data, Mapping):
+        message = "parent must be an object or null"
+        raise TypeError(message)
+    bone_data = data.get("bone")
+    if not isinstance(bone_data, Mapping):
+        message = "parent.bone must be an object"
+        raise TypeError(message)
+    bone = NormalizedBone.from_json(bone_data)
+    human_bone_name_str = data.get("humanBoneName")
+    if not isinstance(human_bone_name_str, str):
+        message = "parent.humanBoneName must be a string"
+        raise TypeError(message)
+    human_bone_specification = HumanBoneSpecifications.from_name_str(
+        human_bone_name_str
+    )
+    if human_bone_specification is None:
+        message = f"Unknown human bone name: {human_bone_name_str}"
+        raise ValueError(message)
+    return (bone, human_bone_specification)
+
+
+def _search_branch_from_json(data: Json) -> SearchBranch:
+    if not isinstance(data, Mapping):
+        message = "SearchBranch must be an object"
+        raise TypeError(message)
+    type_name = data.get("type_name")
+    if type_name == AssignedSearchBranch.__name__:
+        return AssignedSearchBranch.from_json(data)
+
+    if type_name == UnassignedSearchBranch.__name__:
+        return UnassignedSearchBranch.from_json(data)
+
+    message = (
+        "SearchBranch.type_name must be 'AssignedSearchBranch'"
+        + " or 'UnassignedSearchBranch'"
+    )
+    raise ValueError(message)
+
+
+@dataclass(frozen=True)
+class UnassignedSearchBranch:
+    children: tuple[SearchBranch, ...] = ()
+
+    def to_json(self) -> Json:
+        return {
+            "type_name": type(self).__name__,
+            "children": [child.to_json() for child in self.children],
+        }
+
+    @classmethod
+    def from_json(cls, data: Mapping[str, Json]) -> "UnassignedSearchBranch":
+        children_data = data.get("children")
+        if children_data is None:
+            children_data = []
+        if not isinstance(children_data, list):
+            message = "UnassignedSearchBranch.children must be a list"
+            raise TypeError(message)
+        return UnassignedSearchBranch(
+            children=tuple(_search_branch_from_json(child) for child in children_data)
+        )
+
+    @cached_property
+    def score(self) -> float:
+        return sum((child.score for child in self.children), start=0.0)
+
+    @cached_property
+    def recursive_mappings(self) -> Mapping[HumanBoneSpecification, NormalizedBone]:
+        result: dict[HumanBoneSpecification, NormalizedBone] = {}
+        for child in self.children:
+            result.update(child.recursive_mappings)
+        return result
+
+    def __lt__(self, other: "SearchBranch") -> bool:
+        return self.score < other.score
+
+
+@dataclass(frozen=True)
+class AssignedSearchBranch:
+    depth: int
+    bone: NormalizedBone
+    human_bone_specification: HumanBoneSpecification
+    parent: Optional[tuple[NormalizedBone, HumanBoneSpecification]]
+    children: tuple[SearchBranch, ...]
+
+    def to_json(self) -> Json:
+        parent = self.parent
+        if parent is None:
+            parent_json: Json = None
+        else:
+            parent_bone, parent_spec = parent
+            parent_json = {
+                "bone": parent_bone.to_json(),
+                "humanBoneName": parent_spec.name.value,
+            }
+        return {
+            "type_name": type(self).__name__,
+            "depth": self.depth,
+            "bone": self.bone.to_json(),
+            "humanBoneName": self.human_bone_specification.name.value,
+            "parent": parent_json,
+            "children": [child.to_json() for child in self.children],
+        }
+
+    @classmethod
+    def from_json(cls, data: Mapping[str, Json]) -> "AssignedSearchBranch":
+        depth = data.get("depth")
+        if not isinstance(depth, int):
+            message = "AssignedSearchBranch.depth must be an int"
+            raise TypeError(message)
+        bone_data = data.get("bone")
+        if not isinstance(bone_data, Mapping):
+            message = "AssignedSearchBranch.bone must be an object"
+            raise TypeError(message)
+        bone = NormalizedBone.from_json(bone_data)
+        human_bone_name_str = data.get("humanBoneName")
+        if not isinstance(human_bone_name_str, str):
+            message = "AssignedSearchBranch.humanBoneName must be a string"
+            raise TypeError(message)
+        human_bone_specification = HumanBoneSpecifications.from_name_str(
+            human_bone_name_str
+        )
+        if human_bone_specification is None:
+            message = f"Unknown human bone name: {human_bone_name_str}"
+            raise ValueError(message)
+        parent = _parent_from_json(data.get("parent"))
+        children_data = data.get("children")
+        if children_data is None:
+            children_data = []
+        if not isinstance(children_data, list):
+            message = "AssignedSearchBranch.children must be a list"
+            raise TypeError(message)
+        children = tuple(_search_branch_from_json(child) for child in children_data)
+        return AssignedSearchBranch(
+            depth=depth,
+            bone=bone,
+            human_bone_specification=human_bone_specification,
+            parent=parent,
+            children=children,
+        )
+
+    @cached_property
+    def score(self) -> float:
+        total_score = sum((child.score for child in self.children), start=0.0)
+
+        total_score -= math.sqrt(self.depth)
+
+        if self.human_bone_specification.requirement:
+            total_score += 1000
+        else:
+            total_score += 20
+
+        human_bone_name = self.human_bone_specification.name
+
+        if self.bone.z > 0:
+            total_score += 5
+
+        if human_bone_name in CENTRAL_BONE_NAMES:
+            total_score += self.bone.z - abs(self.bone.x)
+        elif human_bone_name in LEFT_ARM_BONE_NAMES:
+            total_score += self.bone.x / 2
+        elif human_bone_name in LEFT_LEG_BONE_NAMES:
+            total_score += 0.2 if self.bone.x > 0 else 0
+        elif human_bone_name in RIGHT_ARM_BONE_NAMES:
+            total_score += -self.bone.x / 2
+        elif human_bone_name in RIGHT_LEG_BONE_NAMES:
+            total_score += 0.2 if self.bone.x < 0 else 0
+
+        parent = self.parent
+        if parent is None:
+            return total_score
+
+        parent_bone, parent_human_bone_specification = parent
+        human_bone_name = self.human_bone_specification.name
+        parent_human_bone_name = parent_human_bone_specification.name
+        if human_bone_name == HumanBoneName.SPINE:
+            if (
+                parent_human_bone_name == HumanBoneName.HIPS
+                and parent_bone.z < self.bone.z
+            ):
+                total_score += 5
+        elif human_bone_name == HumanBoneName.CHEST:
+            if (
+                parent_human_bone_name in HIPS_TO_SPINE_BONE_NAMES
+                and parent_bone.z < self.bone.z
+            ):
+                total_score += 5
+        elif human_bone_name == HumanBoneName.UPPER_CHEST:
+            if (
+                parent_human_bone_name == HumanBoneName.CHEST
+                and parent_bone.z < self.bone.z
+            ):
+                total_score += 5
+        elif human_bone_name == HumanBoneName.NECK:
+            if (
+                parent_human_bone_name in SPINE_TO_UPPER_CHEST_BONE_NAMES
+                and parent_bone.z < self.bone.z
+            ):
+                total_score += 5
+        elif human_bone_name == HumanBoneName.HEAD:
+            if (
+                parent_human_bone_name in SPINE_TO_NECK_NAMES
+                and parent_bone.z < self.bone.z
+            ):
+                total_score += 5
+        else:
+            pass
+
+        return total_score
+
+    @cached_property
+    def recursive_mappings(self) -> Mapping[HumanBoneSpecification, NormalizedBone]:
+        result: dict[HumanBoneSpecification, NormalizedBone] = {
+            self.human_bone_specification: self.bone
+        }
+        for child in self.children:
+            result.update(child.recursive_mappings)
+        return result
+
+    def __lt__(self, other: "SearchBranch") -> bool:
+        return self.score < other.score
+
+
+SEARCH_STOP_BRANCH: Final[SearchBranch] = UnassignedSearchBranch()
+
+
+@dataclass(frozen=True)
+class SearchContext:
+    max_search_count: int
+    require_requirement: int
+    only_requirement: bool
+    counter: Counter = field(default_factory=Counter)
+    cache: dict[
+        tuple[tuple[str, ...], tuple[str, ...], Optional[tuple[str, str]], int],
+        SearchBranch,
+    ] = field(
+        default_factory=dict[
+            tuple[tuple[str, ...], tuple[str, ...], Optional[tuple[str, str]], int],
+            SearchBranch,
+        ]
+    )
+    max_recursion_depth: int = 50
+
+
+DEFAULT_MAX_SEARCH_COUNT: Final[int] = 1000000
+
+
+def create_structure_based_mapping(
+    armature: Object,
+    human_bone_specification: HumanBoneSpecification = HumanBoneSpecifications.HIPS,
+    max_search_count: int = DEFAULT_MAX_SEARCH_COUNT,
+) -> Mapping[str, HumanBoneSpecification]:
+    if not isinstance(armature_data := armature.data, Armature):
+        message = f"{type(armature.data)} is not an Armature"
+        raise TypeError(message)
+
+    pose_position = armature_data.pose_position
+    armature_data.pose_position = "REST"
+
+    translations = [
+        (armature.matrix_world @ bone.matrix_local).translation
+        for bone in armature_data.bones
+    ]
+    max_x = max((abs(translation.x) for translation in translations), default=0)
+    max_y = max((abs(translation.y) for translation in translations), default=0)
+    max_z = max((abs(translation.z) for translation in translations), default=0)
+
+    root_bones = [
+        NormalizedBone.create(
+            armature.matrix_world,
+            max_x,
+            max_y,
+            max_z,
+            bone,
+        )
+        for bone in armature_data.bones
+        if bone.parent is None
+    ]
+    armature_data.pose_position = pose_position
+    counter = Counter()
+    max_result: SearchBranch = SEARCH_STOP_BRANCH
+
+    for require_requirement, only_requirement in ((1, True), (0, False)):
+        search_context = SearchContext(
+            require_requirement=require_requirement,
+            only_requirement=only_requirement,
+            max_search_count=max_search_count,
+            counter=counter,
+        )
+        for skip_count in range(1000):
+            if counter.count > search_context.max_search_count:
+                break
+            changed = False
+            for root_bone in root_bones:
+                if counter.count > search_context.max_search_count:
+                    break
+                if (
+                    root_bone.recursive_len
+                    < skip_count
+                    + HumanBoneSpecifications.HIPS.recursive_requirement_len
+                    * require_requirement
+                ):
+                    break
+                result = _search_structure_based_mapping_step(
+                    root_bone,
+                    (human_bone_specification,),
+                    None,
+                    1,
+                    skip_count,
+                    search_context,
+                )
+                if max_result < result:
+                    max_result = result
+                    changed = True
+
+            if not changed:
+                continue
+
+            _logger.warning("Structure search context:")
+            _logger.warning("  Search Count Limit: %d", search_context.max_search_count)
+            _logger.warning("  Recursion Limit: %d", search_context.max_recursion_depth)
+            _logger.warning("  Skip Count: %d", skip_count)
+            _logger.warning(
+                "  Require Requirement: %d", search_context.require_requirement
+            )
+            _logger.warning("  Only Requirement: %d", search_context.only_requirement)
+            _logger.warning(
+                "Structure search: Total searches: %d, Cache hits: %d",
+                counter.count,
+                counter.cache_hit,
+            )
+            _logger.warning(
+                "Generated mappings:\n  %s",
+                "\n  ".join(
+                    f"{spec.name.value}: {bone.name}"
+                    for spec, bone in max_result.recursive_mappings.items()
+                ),
+            )
+
+    _logger.warning(
+        "Structure search completed:"
+        " Total searches: %d (soft limit: %d), Cache hits: %d",
+        counter.count,
+        max_search_count,
+        counter.cache_hit,
+    )
+    mapping = max_result.recursive_mappings
+    return {spec.name: bone for bone, spec in mapping.items()}
+
+
+def _search_structure_based_mapping_step(
+    bone: NormalizedBone,
+    human_bone_specifications: tuple[HumanBoneSpecification, ...],
+    parent: Optional[tuple[NormalizedBone, HumanBoneSpecification]],
+    depth: int,
+    skip_count: int,
+    search_context: SearchContext,
+) -> SearchBranch:
+    if skip_count < 0 or bone.recursive_len < skip_count:
+        return SEARCH_STOP_BRANCH
+
+    if search_context.require_requirement:
+        recursive_requirement_len = sum(
+            human_bone_specification.recursive_requirement_len
+            for human_bone_specification in human_bone_specifications
+        )
+        if bone.recursive_len < skip_count + recursive_requirement_len:
+            return SEARCH_STOP_BRANCH
+
+    if depth > search_context.max_recursion_depth:
+        return SEARCH_STOP_BRANCH
+
+    counter = search_context.counter
+    counter.count += 1
+    if counter.count > search_context.max_search_count:
+        return SEARCH_STOP_BRANCH
+
+    cache_key = (
+        (bone.name,),
+        tuple(spec.name.value for spec in human_bone_specifications),
+        None
+        if parent is None
+        else (
+            parent[0].name,
+            parent[1].name.value,
+        ),
+        skip_count,
+    )
+    cache = search_context.cache
+    if (search_branchs := cache.get(cache_key)) is not None:
+        counter.cache_hit += 1
+        return search_branchs
+
+    result = SEARCH_STOP_BRANCH
+
+    if len(human_bone_specifications) == 1:
+        human_bone_specification = human_bone_specifications[0]
+        human_bone_specification_children = (
+            human_bone_specification.required_children
+            if search_context.only_requirement
+            else human_bone_specification.children
+        )
+        if bone.children and human_bone_specification_children:
+            children = (
+                _search_structure_based_mapping_product(
+                    bone.children,
+                    human_bone_specification_children,
+                    (bone, human_bone_specification),
+                    depth + 1,
+                    skip_count,
+                    search_context,
+                ),
+            )
+        else:
+            children = ()
+
+        result = max(
+            result,
+            AssignedSearchBranch(
+                depth=depth + 1,
+                bone=bone,
+                human_bone_specification=human_bone_specification,
+                parent=parent,
+                children=children,
+            ),
+        )
+
+    if skip_count > 0:
+        result = max(
+            result,
+            _search_structure_based_mapping_product(
+                bone.children,
+                human_bone_specifications,
+                parent,
+                depth + 1,
+                skip_count - 1,
+                search_context,
+            ),
+        )
+
+    cache[cache_key] = result
+    return result
+
+
+def _search_structure_based_mapping_product(
+    bones: tuple[NormalizedBone, ...],
+    human_bone_specifications: tuple[HumanBoneSpecification, ...],
+    parent: Optional[tuple[NormalizedBone, HumanBoneSpecification]],
+    depth: int,
+    skip_count: int,
+    search_context: SearchContext,
+) -> SearchBranch:
+    if not bones or not human_bone_specifications:
+        return SEARCH_STOP_BRANCH
+
+    if depth > search_context.max_recursion_depth:
+        return SEARCH_STOP_BRANCH
+
+    if skip_count < 0 or sum(bone.recursive_len for bone in bones) < skip_count:
+        return SEARCH_STOP_BRANCH
+
+    counter = search_context.counter
+    counter.count += 1
+    if counter.count > search_context.max_search_count:
+        return SEARCH_STOP_BRANCH
+
+    cache = search_context.cache
+    cache_key = (
+        tuple(sorted(bone.name for bone in bones)),
+        tuple(spec.name.value for spec in human_bone_specifications),
+        None
+        if parent is None
+        else (
+            parent[0].name,
+            parent[1].name.value,
+        ),
+        skip_count,
+    )
+    if (search_branchs := cache.get(cache_key)) is not None:
+        counter.cache_hit += 1
+        return search_branchs
+
+    result = SEARCH_STOP_BRANCH
+
+    for skipping_human_bone_specification in human_bone_specifications:
+        if skipping_human_bone_specification.requirement:
+            continue
+        skipping_children = sum(
+            tuple(
+                skipping_human_bone_specification.children
+                if human_bone_specification == skipping_human_bone_specification
+                else (human_bone_specification,)
+                for human_bone_specification in human_bone_specifications
+            ),
+            (),
+        )
+        if skipping_children:
+            result = max(
+                result,
+                _search_structure_based_mapping_product(
+                    bones,
+                    skipping_children,
+                    parent,
+                    depth + 1,
+                    skip_count,
+                    search_context,
+                ),
+            )
+
+    assignment_combination_len = min(len(human_bone_specifications), len(bones))
+    if (
+        not search_context.require_requirement
+        or sum(int(s.requirement) for s in human_bone_specifications)
+        <= assignment_combination_len
+    ):
+        spec_tuples = tuple(
+            (human_bone_specification,)
+            for human_bone_specification in human_bone_specifications
+        )
+        for spec_children, bone_children in itertools.product(
+            itertools.combinations(spec_tuples, assignment_combination_len),
+            itertools.combinations(bones, assignment_combination_len),
+        ):
+            for spec_children_permutation in itertools.permutations(spec_children):
+                for bone_packings in _iter_skip_counts_and_bone_combinations(
+                    skip_count,
+                    tuple(
+                        zip(
+                            bone_children,
+                            (
+                                sum(s.recursive_requirement_len for s in ss)
+                                * search_context.require_requirement
+                                for ss in spec_children_permutation
+                            ),
+                        )
+                    ),
+                ):
+                    children = tuple(
+                        _search_structure_based_mapping_step(
+                            bone_child,
+                            spec_child,
+                            parent,
+                            depth + 1,
+                            child_skip_count,
+                            search_context,
+                        )
+                        for spec_child, (bone_child, child_skip_count) in zip(
+                            spec_children_permutation, bone_packings
+                        )
+                    )
+                    result = max(result, UnassignedSearchBranch(children=children))
+                    if counter.count > search_context.max_search_count:
+                        break
+
+    for special_group in (
+        (
+            HumanBoneSpecifications.LEFT_UPPER_LEG,
+            HumanBoneSpecifications.RIGHT_UPPER_LEG,
+        ),
+        (
+            HumanBoneSpecifications.LEFT_UPPER_ARM,
+            HumanBoneSpecifications.RIGHT_UPPER_ARM,
+        ),
+        (
+            HumanBoneSpecifications.LEFT_EYE,
+            HumanBoneSpecifications.RIGHT_EYE,
+        ),
+        # (HumanBoneSpecifications.LEFT_THUMB_METACARPAL,),
+        # (HumanBoneSpecifications.RIGHT_THUMB_METACARPAL,),
+    ):
+        if not all(s in human_bone_specifications for s in special_group):
+            continue
+
+        remaining_group = [
+            s for s in human_bone_specifications if s not in special_group
+        ]
+
+        if remaining_group:
+            groups = (
+                special_group,
+                *tuple((remaining,) for remaining in remaining_group),
+            )
+        else:
+            groups = (special_group,)
+
+        special_assignment_combination_len = min(len(groups), len(bones))
+        if (
+            search_context.require_requirement
+            and special_assignment_combination_len
+            < sum(int(any(s.requirement for s in group)) for group in groups)
+        ):
+            break
+
+        for spec_children, bone_children in itertools.product(
+            itertools.combinations(groups, special_assignment_combination_len),
+            itertools.combinations(bones, special_assignment_combination_len),
+        ):
+            for spec_children_permutation in itertools.permutations(spec_children):
+                for bone_packings in _iter_skip_counts_and_bone_combinations(
+                    skip_count,
+                    tuple(
+                        zip(
+                            bone_children,
+                            (
+                                sum(s.recursive_requirement_len for s in ss)
+                                * search_context.require_requirement
+                                for ss in spec_children_permutation
+                            ),
+                        )
+                    ),
+                ):
+                    children = tuple(
+                        _search_structure_based_mapping_step(
+                            bone_child,
+                            spec_child,
+                            parent,
+                            depth + 1,
+                            child_skip_count,
+                            search_context,
+                        )
+                        for spec_child, (bone_child, child_skip_count) in zip(
+                            spec_children_permutation, bone_packings
+                        )
+                    )
+                    result = max(result, UnassignedSearchBranch(children=children))
+                    if counter.count > search_context.max_search_count:
+                        break
+        break
+
+    cache[cache_key] = result
+    return result
+
+
+def _backtrack_skip_counts_and_bone_combinations(
+    *,
+    bones: tuple[tuple[NormalizedBone, int], ...],
+    counts: list[int],
+    bone_index: int,
+    remaining: int,
+) -> Iterator[list[tuple[NormalizedBone, int]]]:
+    if bone_index == len(bones):
+        if remaining == 0:
+            yield [(bone, count) for (bone, _), count in zip(bones, counts)]
+        return
+
+    bone, requirement_len = bones[bone_index]
+    max_count = min(bone.recursive_len - requirement_len, remaining)
+    for count in range(max_count, -1, -1):
+        counts.append(count)
+        yield from _backtrack_skip_counts_and_bone_combinations(
+            bones=bones,
+            counts=counts,
+            bone_index=bone_index + 1,
+            remaining=remaining - count,
+        )
+        counts.pop()
+
+
+def _iter_skip_counts_and_bone_combinations(
+    skip_count: int, bones: tuple[tuple[NormalizedBone, int], ...]
+) -> Iterator[list[tuple[NormalizedBone, int]]]:
+    if skip_count < 0:
+        return
+
+    if not bones:
+        if skip_count == 0:
+            yield []
+        return
+
+    total_recursive_len = sum(
+        b.recursive_len - requirement_len for b, requirement_len in bones
+    )
+    if total_recursive_len < skip_count:
+        return
+
+    yield from _backtrack_skip_counts_and_bone_combinations(
+        bones=bones,
+        counts=[],
+        bone_index=0,
+        remaining=skip_count,
+    )