@UaRuairc has signed the CLA in opendatalab/MinerU#4654

.github/workflows/cla.yml

@@ -18,9 +18,9 @@ jobs:
     steps:
       - name: "CLA Assistant"
         if: (github.event.comment.body == 'recheck' || github.event.comment.body == 'I have read the CLA Document and I hereby sign the CLA') || github.event_name == 'pull_request_target'
-        uses: contributor-assistant/github-action@v2.6.1
+        uses: contributor-assistant/github-action@v2.5.0
         env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN  }}
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # the below token should have repo scope and must be manually added by you in the repository's secret
           # This token is required only if you have configured to store the signatures in a remote repository/organization
           PERSONAL_ACCESS_TOKEN: ${{ secrets.RELEASE_TOKEN }}
@@ -28,7 +28,7 @@ jobs:
           path-to-signatures: 'signatures/version1/cla.json'
           path-to-document: 'https://github.com/opendatalab/MinerU/blob/master/MinerU_CLA.md' # e.g. a CLA or a DCO document
           # branch should not be protected
-          branch: 'cla'
+          branch: 'master'
           allowlist: myhloli,dt-yy,Focusshang,renpengli01,icecraft,drunkpig,wangbinDL,qiangqiang199,GDDGCZ518,papayalove,conghui,quyuan,LollipopsAndWine,Sidney233
 
          # the followings are the optional inputs - If the optional inputs are not given, then default values will be taken

README.md

@@ -43,10 +43,7 @@
 </div>
 
 # Changelog
-- 2025/07/23 2.1.4 Released
-  - Bug Fixes
-    - Fixed the issue of excessive memory consumption during the `MFR` step in the `pipeline` backend under certain scenarios #2771
-    - Fixed the inaccurate matching between `image`/`table` and `caption`/`footnote` under certain conditions #3129
+
 - 2025/07/16 2.1.1 Released
   - Bug fixes
     - Fixed text block content loss issue that could occur in certain `pipeline` scenarios #3005

README_zh-CN.md

@@ -43,10 +43,6 @@
 </div>
 
 # 更新记录
-- 2025/07/23 2.1.4发布
-  - bug修复
-    - 修复`pipeline`后端中`MFR`步骤在某些情况下显存消耗过大的问题 #2771
-    - 修复某些情况下`image`/`table`与`caption`/`footnote`匹配不准确的问题 #3129
 - 2025/07/16 2.1.1发布
   - bug修复 
     - 修复`pipeline`在某些情况可能发生的文本块内容丢失问题 #3005

mineru/backend/pipeline/batch_analyze.py

@@ -12,7 +12,6 @@ from ...utils.ocr_utils import get_adjusted_mfdetrec_res, get_ocr_result_list, O
 YOLO_LAYOUT_BASE_BATCH_SIZE = 8
 MFD_BASE_BATCH_SIZE = 1
 MFR_BASE_BATCH_SIZE = 16
-OCR_DET_BASE_BATCH_SIZE = 16
 
 
 class BatchAnalyze:
@@ -171,9 +170,9 @@ class BatchAnalyze:
                         batch_images.append(padded_img)
 
                     # 批处理检测
-                    det_batch_size = min(len(batch_images), self.batch_ratio * OCR_DET_BASE_BATCH_SIZE)  # 增加批处理大小
-                    # logger.debug(f"OCR-det batch: {det_batch_size} images, target size: {target_h}x{target_w}")
-                    batch_results = ocr_model.text_detector.batch_predict(batch_images, det_batch_size)
+                    batch_size = min(len(batch_images), self.batch_ratio * 16)  # 增加批处理大小
+                    # logger.debug(f"OCR-det batch: {batch_size} images, target size: {target_h}x{target_w}")
+                    batch_results = ocr_model.text_detector.batch_predict(batch_images, batch_size)
 
                     # 处理批处理结果
                     for i, (crop_info, (dt_boxes, elapse)) in enumerate(zip(group_crops, batch_results)):

mineru/backend/pipeline/pipeline_analyze.py

@@ -74,10 +74,10 @@ def doc_analyze(
         table_enable=True,
 ):
     """
-    适当调大MIN_BATCH_INFERENCE_SIZE可以提高性能，更大的 MIN_BATCH_INFERENCE_SIZE会消耗更多内存，
-    可通过环境变量MINERU_MIN_BATCH_INFERENCE_SIZE设置，默认值为384。
+    适当调大MIN_BATCH_INFERENCE_SIZE可以提高性能，可能会增加显存使用量，
+    可通过环境变量MINERU_MIN_BATCH_INFERENCE_SIZE设置，默认值为128。
     """
-    min_batch_inference_size = int(os.environ.get('MINERU_MIN_BATCH_INFERENCE_SIZE', 384))
+    min_batch_inference_size = int(os.environ.get('MINERU_MIN_BATCH_INFERENCE_SIZE', 128))
 
     # 收集所有页面信息
     all_pages_info = []  # 存储(dataset_index, page_index, img, ocr, lang, width, height)

mineru/backend/pipeline/pipeline_magic_model.py

@@ -1,6 +1,5 @@
-from mineru.utils.boxbase import bbox_relative_pos, calculate_iou, bbox_distance, get_minbox_if_overlap_by_ratio
+from mineru.utils.boxbase import bbox_relative_pos, calculate_iou, bbox_distance, is_in, get_minbox_if_overlap_by_ratio
 from mineru.utils.enum_class import CategoryId, ContentType
-from mineru.utils.magic_model_utils import tie_up_category_by_distance_v3, reduct_overlap
 
 
 class MagicModel:
@@ -209,39 +208,168 @@ class MagicModel:
 
         return bbox_distance(bbox1, bbox2)
 
-    def __tie_up_category_by_distance_v3(self, subject_category_id, object_category_id):
-        # 定义获取主体和客体对象的函数
-        def get_subjects():
-            return reduct_overlap(
-                list(
-                    map(
-                        lambda x: {'bbox': x['bbox'], 'score': x['score']},
-                        filter(
-                            lambda x: x['category_id'] == subject_category_id,
-                            self.__page_model_info['layout_dets'],
-                        ),
-                    )
+    def __reduct_overlap(self, bboxes):
+        N = len(bboxes)
+        keep = [True] * N
+        for i in range(N):
+            for j in range(N):
+                if i == j:
+                    continue
+                if is_in(bboxes[i]['bbox'], bboxes[j]['bbox']):
+                    keep[i] = False
+        return [bboxes[i] for i in range(N) if keep[i]]
+
+    def __tie_up_category_by_distance_v3(
+        self,
+        subject_category_id: int,
+        object_category_id: int,
+    ):
+        subjects = self.__reduct_overlap(
+            list(
+                map(
+                    lambda x: {'bbox': x['bbox'], 'score': x['score']},
+                    filter(
+                        lambda x: x['category_id'] == subject_category_id,
+                        self.__page_model_info['layout_dets'],
+                    ),
                 )
             )
-
-        def get_objects():
-            return reduct_overlap(
-                list(
-                    map(
-                        lambda x: {'bbox': x['bbox'], 'score': x['score']},
-                        filter(
-                            lambda x: x['category_id'] == object_category_id,
-                            self.__page_model_info['layout_dets'],
-                        ),
-                    )
-                )
-            )
-
-        # 调用通用方法
-        return tie_up_category_by_distance_v3(
-            get_subjects,
-            get_objects
         )
+        objects = self.__reduct_overlap(
+            list(
+                map(
+                    lambda x: {'bbox': x['bbox'], 'score': x['score']},
+                    filter(
+                        lambda x: x['category_id'] == object_category_id,
+                        self.__page_model_info['layout_dets'],
+                    ),
+                )
+            )
+        )
+
+        ret = []
+        N, M = len(subjects), len(objects)
+        subjects.sort(key=lambda x: x['bbox'][0] ** 2 + x['bbox'][1] ** 2)
+        objects.sort(key=lambda x: x['bbox'][0] ** 2 + x['bbox'][1] ** 2)
+
+        OBJ_IDX_OFFSET = 10000
+        SUB_BIT_KIND, OBJ_BIT_KIND = 0, 1
+
+        all_boxes_with_idx = [(i, SUB_BIT_KIND, sub['bbox'][0], sub['bbox'][1]) for i, sub in enumerate(subjects)] + [(i + OBJ_IDX_OFFSET , OBJ_BIT_KIND, obj['bbox'][0], obj['bbox'][1]) for i, obj in enumerate(objects)]
+        seen_idx = set()
+        seen_sub_idx = set()
+
+        while N > len(seen_sub_idx):
+            candidates = []
+            for idx, kind, x0, y0 in all_boxes_with_idx:
+                if idx in seen_idx:
+                    continue
+                candidates.append((idx, kind, x0, y0))
+
+            if len(candidates) == 0:
+                break
+            left_x = min([v[2] for v in candidates])
+            top_y =  min([v[3] for v in candidates])
+
+            candidates.sort(key=lambda x: (x[2]-left_x) ** 2 + (x[3] - top_y) ** 2)
+
+
+            fst_idx, fst_kind, left_x, top_y = candidates[0]
+            candidates.sort(key=lambda x: (x[2] - left_x) ** 2 + (x[3] - top_y)**2)
+            nxt = None
+
+            for i in range(1, len(candidates)):
+                if candidates[i][1] ^ fst_kind == 1:
+                    nxt = candidates[i]
+                    break
+            if nxt is None:
+                break
+
+            if fst_kind == SUB_BIT_KIND:
+                sub_idx, obj_idx = fst_idx, nxt[0] - OBJ_IDX_OFFSET
+
+            else:
+                sub_idx, obj_idx = nxt[0], fst_idx - OBJ_IDX_OFFSET
+
+            pair_dis = bbox_distance(subjects[sub_idx]['bbox'], objects[obj_idx]['bbox'])
+            nearest_dis = float('inf')
+            for i in range(N):
+                if i in seen_idx or i == sub_idx:continue
+                nearest_dis = min(nearest_dis, bbox_distance(subjects[i]['bbox'], objects[obj_idx]['bbox']))
+
+            if pair_dis >= 3*nearest_dis:
+                seen_idx.add(sub_idx)
+                continue
+
+            seen_idx.add(sub_idx)
+            seen_idx.add(obj_idx + OBJ_IDX_OFFSET)
+            seen_sub_idx.add(sub_idx)
+
+            ret.append(
+                {
+                    'sub_bbox': {
+                        'bbox': subjects[sub_idx]['bbox'],
+                        'score': subjects[sub_idx]['score'],
+                    },
+                    'obj_bboxes': [
+                        {'score': objects[obj_idx]['score'], 'bbox': objects[obj_idx]['bbox']}
+                    ],
+                    'sub_idx': sub_idx,
+                }
+            )
+
+        for i in range(len(objects)):
+            j = i + OBJ_IDX_OFFSET
+            if j in seen_idx:
+                continue
+            seen_idx.add(j)
+            nearest_dis, nearest_sub_idx = float('inf'), -1
+            for k in range(len(subjects)):
+                dis = bbox_distance(objects[i]['bbox'], subjects[k]['bbox'])
+                if dis < nearest_dis:
+                    nearest_dis = dis
+                    nearest_sub_idx = k
+
+            for k in range(len(subjects)):
+                if k != nearest_sub_idx: continue
+                if k in seen_sub_idx:
+                    for kk in range(len(ret)):
+                        if ret[kk]['sub_idx'] == k:
+                            ret[kk]['obj_bboxes'].append({'score': objects[i]['score'], 'bbox': objects[i]['bbox']})
+                            break
+                else:
+                    ret.append(
+                        {
+                            'sub_bbox': {
+                                'bbox': subjects[k]['bbox'],
+                                'score': subjects[k]['score'],
+                            },
+                            'obj_bboxes': [
+                                {'score': objects[i]['score'], 'bbox': objects[i]['bbox']}
+                            ],
+                            'sub_idx': k,
+                        }
+                    )
+                seen_sub_idx.add(k)
+                seen_idx.add(k)
+
+
+        for i in range(len(subjects)):
+            if i in seen_sub_idx:
+                continue
+            ret.append(
+                {
+                    'sub_bbox': {
+                        'bbox': subjects[i]['bbox'],
+                        'score': subjects[i]['score'],
+                    },
+                    'obj_bboxes': [],
+                    'sub_idx': i,
+                }
+            )
+
+
+        return ret
 
     def get_imgs(self):
         with_captions = self.__tie_up_category_by_distance_v3(

mineru/backend/vlm/vlm_magic_model.py

@@ -3,10 +3,10 @@ from typing import Literal
 
 from loguru import logger
 
+from mineru.utils.boxbase import bbox_distance, is_in
 from mineru.utils.enum_class import ContentType, BlockType, SplitFlag
 from mineru.backend.vlm.vlm_middle_json_mkcontent import merge_para_with_text
 from mineru.utils.format_utils import convert_otsl_to_html
-from mineru.utils.magic_model_utils import reduct_overlap, tie_up_category_by_distance_v3
 
 
 class MagicModel:
@@ -251,39 +251,175 @@ def latex_fix(latex):
     return latex
 
 
-def __tie_up_category_by_distance_v3(blocks, subject_block_type, object_block_type):
-    # 定义获取主体和客体对象的函数
-    def get_subjects():
-        return reduct_overlap(
-            list(
-                map(
-                    lambda x: {"bbox": x["bbox"], "lines": x["lines"], "index": x["index"]},
-                    filter(
-                        lambda x: x["type"] == subject_block_type,
-                        blocks,
-                    ),
-                )
+def __reduct_overlap(bboxes):
+    N = len(bboxes)
+    keep = [True] * N
+    for i in range(N):
+        for j in range(N):
+            if i == j:
+                continue
+            if is_in(bboxes[i]["bbox"], bboxes[j]["bbox"]):
+                keep[i] = False
+    return [bboxes[i] for i in range(N) if keep[i]]
+
+
+def __tie_up_category_by_distance_v3(
+    blocks: list,
+    subject_block_type: str,
+    object_block_type: str,
+):
+    subjects = __reduct_overlap(
+        list(
+            map(
+                lambda x: {"bbox": x["bbox"], "lines": x["lines"], "index": x["index"]},
+                filter(
+                    lambda x: x["type"] == subject_block_type,
+                    blocks,
+                ),
             )
         )
-
-    def get_objects():
-        return reduct_overlap(
-            list(
-                map(
-                    lambda x: {"bbox": x["bbox"], "lines": x["lines"], "index": x["index"]},
-                    filter(
-                        lambda x: x["type"] == object_block_type,
-                        blocks,
-                    ),
-                )
-            )
-        )
-
-    # 调用通用方法
-    return tie_up_category_by_distance_v3(
-        get_subjects,
-        get_objects
     )
+    objects = __reduct_overlap(
+        list(
+            map(
+                lambda x: {"bbox": x["bbox"], "lines": x["lines"], "index": x["index"]},
+                filter(
+                    lambda x: x["type"] == object_block_type,
+                    blocks,
+                ),
+            )
+        )
+    )
+
+    ret = []
+    N, M = len(subjects), len(objects)
+    subjects.sort(key=lambda x: x["bbox"][0] ** 2 + x["bbox"][1] ** 2)
+    objects.sort(key=lambda x: x["bbox"][0] ** 2 + x["bbox"][1] ** 2)
+
+    OBJ_IDX_OFFSET = 10000
+    SUB_BIT_KIND, OBJ_BIT_KIND = 0, 1
+
+    all_boxes_with_idx = [(i, SUB_BIT_KIND, sub["bbox"][0], sub["bbox"][1]) for i, sub in enumerate(subjects)] + [
+        (i + OBJ_IDX_OFFSET, OBJ_BIT_KIND, obj["bbox"][0], obj["bbox"][1]) for i, obj in enumerate(objects)
+    ]
+    seen_idx = set()
+    seen_sub_idx = set()
+
+    while N > len(seen_sub_idx):
+        candidates = []
+        for idx, kind, x0, y0 in all_boxes_with_idx:
+            if idx in seen_idx:
+                continue
+            candidates.append((idx, kind, x0, y0))
+
+        if len(candidates) == 0:
+            break
+        left_x = min([v[2] for v in candidates])
+        top_y = min([v[3] for v in candidates])
+
+        candidates.sort(key=lambda x: (x[2] - left_x) ** 2 + (x[3] - top_y) ** 2)
+
+        fst_idx, fst_kind, left_x, top_y = candidates[0]
+        candidates.sort(key=lambda x: (x[2] - left_x) ** 2 + (x[3] - top_y) ** 2)
+        nxt = None
+
+        for i in range(1, len(candidates)):
+            if candidates[i][1] ^ fst_kind == 1:
+                nxt = candidates[i]
+                break
+        if nxt is None:
+            break
+
+        if fst_kind == SUB_BIT_KIND:
+            sub_idx, obj_idx = fst_idx, nxt[0] - OBJ_IDX_OFFSET
+
+        else:
+            sub_idx, obj_idx = nxt[0], fst_idx - OBJ_IDX_OFFSET
+
+        pair_dis = bbox_distance(subjects[sub_idx]["bbox"], objects[obj_idx]["bbox"])
+        nearest_dis = float("inf")
+        for i in range(N):
+            if i in seen_idx or i == sub_idx:
+                continue
+            nearest_dis = min(nearest_dis, bbox_distance(subjects[i]["bbox"], objects[obj_idx]["bbox"]))
+
+        if pair_dis >= 3 * nearest_dis:
+            seen_idx.add(sub_idx)
+            continue
+
+        seen_idx.add(sub_idx)
+        seen_idx.add(obj_idx + OBJ_IDX_OFFSET)
+        seen_sub_idx.add(sub_idx)
+
+        ret.append(
+            {
+                "sub_bbox": {
+                    "bbox": subjects[sub_idx]["bbox"],
+                    "lines": subjects[sub_idx]["lines"],
+                    "index": subjects[sub_idx]["index"],
+                },
+                "obj_bboxes": [
+                    {"bbox": objects[obj_idx]["bbox"], "lines": objects[obj_idx]["lines"], "index": objects[obj_idx]["index"]}
+                ],
+                "sub_idx": sub_idx,
+            }
+        )
+
+    for i in range(len(objects)):
+        j = i + OBJ_IDX_OFFSET
+        if j in seen_idx:
+            continue
+        seen_idx.add(j)
+        nearest_dis, nearest_sub_idx = float("inf"), -1
+        for k in range(len(subjects)):
+            dis = bbox_distance(objects[i]["bbox"], subjects[k]["bbox"])
+            if dis < nearest_dis:
+                nearest_dis = dis
+                nearest_sub_idx = k
+
+        for k in range(len(subjects)):
+            if k != nearest_sub_idx:
+                continue
+            if k in seen_sub_idx:
+                for kk in range(len(ret)):
+                    if ret[kk]["sub_idx"] == k:
+                        ret[kk]["obj_bboxes"].append(
+                            {"bbox": objects[i]["bbox"], "lines": objects[i]["lines"], "index": objects[i]["index"]}
+                        )
+                        break
+            else:
+                ret.append(
+                    {
+                        "sub_bbox": {
+                            "bbox": subjects[k]["bbox"],
+                            "lines": subjects[k]["lines"],
+                            "index": subjects[k]["index"],
+                        },
+                        "obj_bboxes": [
+                            {"bbox": objects[i]["bbox"], "lines": objects[i]["lines"], "index": objects[i]["index"]}
+                        ],
+                        "sub_idx": k,
+                    }
+                )
+            seen_sub_idx.add(k)
+            seen_idx.add(k)
+
+    for i in range(len(subjects)):
+        if i in seen_sub_idx:
+            continue
+        ret.append(
+            {
+                "sub_bbox": {
+                    "bbox": subjects[i]["bbox"],
+                    "lines": subjects[i]["lines"],
+                    "index": subjects[i]["index"],
+                },
+                "obj_bboxes": [],
+                "sub_idx": i,
+            }
+        )
+
+    return ret
 
 
 def get_type_blocks(blocks, block_type: Literal["image", "table"]):

mineru/model/mfr/unimernet/Unimernet.py

@@ -104,10 +104,6 @@ class UnimernetModel(object):
 
         # Create dataset with sorted images
         dataset = MathDataset(sorted_images, transform=self.model.transform)
-
-        # 如果batch_size > len(sorted_images)，则设置为不超过len(sorted_images)的2的幂
-        batch_size = min(batch_size, max(1, 2 ** (len(sorted_images).bit_length() - 1))) if sorted_images else 1
-
         dataloader = DataLoader(dataset, batch_size=batch_size, num_workers=0)
 
         # Process batches and store results
@@ -119,7 +115,7 @@ class UnimernetModel(object):
                 mf_img = mf_img.to(dtype=self.model.dtype)
                 mf_img = mf_img.to(self.device)
                 with torch.no_grad():
-                    output = self.model.generate({"image": mf_img}, batch_size=batch_size)
+                    output = self.model.generate({"image": mf_img})
                 mfr_res.extend(output["fixed_str"])
 
                 # 更新进度条，每次增加batch_size，但要注意最后一个batch可能不足batch_size

mineru/model/mfr/unimernet/unimernet_hf/modeling_unimernet.py

@@ -468,7 +468,7 @@ class UnimernetModel(VisionEncoderDecoderModel):
         ).loss
         return {"loss": loss}
 
-    def generate(self, samples, do_sample: bool = False, temperature: float = 0.2, top_p: float = 0.95, batch_size=64):
+    def generate(self, samples, do_sample: bool = False, temperature: float = 0.2, top_p: float = 0.95):
         pixel_values = samples["image"]
         num_channels = pixel_values.shape[1]
         if num_channels == 1:
@@ -478,13 +478,7 @@ class UnimernetModel(VisionEncoderDecoderModel):
         if do_sample:
             kwargs["temperature"] = temperature
             kwargs["top_p"] = top_p
-
-        if self.tokenizer.tokenizer.model_max_length > 1152:
-            if batch_size <= 32:
-                self.tokenizer.tokenizer.model_max_length = 1152  # 6g
-            else:
-                self.tokenizer.tokenizer.model_max_length = 1344  # 8g
-
+        
         outputs = super().generate(
             pixel_values=pixel_values,
             max_new_tokens=self.tokenizer.tokenizer.model_max_length, # required

mineru/model/ocr/paddleocr2pytorch/pytorch_paddle.py

@@ -88,7 +88,7 @@ class PytorchPaddleOCR(TextSystem):
         kwargs['det_model_path'] = det_model_path
         kwargs['rec_model_path'] = rec_model_path
         kwargs['rec_char_dict_path'] = os.path.join(root_dir, 'pytorchocr', 'utils', 'resources', 'dict', dict_file)
-        kwargs['rec_batch_num'] = 16
+        # kwargs['rec_batch_num'] = 8
 
         kwargs['device'] = device
 

mineru/utils/magic_model_utils.py

@@ -1,168 +0,0 @@
-"""
-包含两个MagicModel类中重复使用的方法和逻辑
-"""
-from typing import List, Dict, Any, Callable
-from mineru.utils.boxbase import bbox_distance, is_in
-
-
-def reduct_overlap(bboxes: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
-    """
-    去除重叠的bbox，保留不被其他bbox包含的bbox
-
-    Args:
-        bboxes: 包含bbox信息的字典列表
-
-    Returns:
-        去重后的bbox列表
-    """
-    N = len(bboxes)
-    keep = [True] * N
-    for i in range(N):
-        for j in range(N):
-            if i == j:
-                continue
-            if is_in(bboxes[i]['bbox'], bboxes[j]['bbox']):
-                keep[i] = False
-    return [bboxes[i] for i in range(N) if keep[i]]
-
-
-def tie_up_category_by_distance_v3(
-        get_subjects_func: Callable,
-        get_objects_func: Callable,
-        extract_subject_func: Callable = None,
-        extract_object_func: Callable = None
-):
-    """
-    通用的类别关联方法，用于将主体对象与客体对象进行关联
-
-    参数:
-        get_subjects_func: 函数，提取主体对象
-        get_objects_func: 函数，提取客体对象
-        extract_subject_func: 函数，自定义提取主体属性（默认使用bbox和其他属性）
-        extract_object_func: 函数，自定义提取客体属性（默认使用bbox和其他属性）
-
-    返回:
-        关联后的对象列表
-    """
-    subjects = get_subjects_func()
-    objects = get_objects_func()
-
-    # 如果没有提供自定义提取函数，使用默认函数
-    if extract_subject_func is None:
-        extract_subject_func = lambda x: x
-    if extract_object_func is None:
-        extract_object_func = lambda x: x
-
-    ret = []
-    N, M = len(subjects), len(objects)
-    subjects.sort(key=lambda x: x["bbox"][0] ** 2 + x["bbox"][1] ** 2)
-    objects.sort(key=lambda x: x["bbox"][0] ** 2 + x["bbox"][1] ** 2)
-
-    OBJ_IDX_OFFSET = 10000
-    SUB_BIT_KIND, OBJ_BIT_KIND = 0, 1
-
-    all_boxes_with_idx = [(i, SUB_BIT_KIND, sub["bbox"][0], sub["bbox"][1]) for i, sub in enumerate(subjects)] + [
-        (i + OBJ_IDX_OFFSET, OBJ_BIT_KIND, obj["bbox"][0], obj["bbox"][1]) for i, obj in enumerate(objects)
-    ]
-    seen_idx = set()
-    seen_sub_idx = set()
-
-    while N > len(seen_sub_idx):
-        candidates = []
-        for idx, kind, x0, y0 in all_boxes_with_idx:
-            if idx in seen_idx:
-                continue
-            candidates.append((idx, kind, x0, y0))
-
-        if len(candidates) == 0:
-            break
-        left_x = min([v[2] for v in candidates])
-        top_y = min([v[3] for v in candidates])
-
-        candidates.sort(key=lambda x: (x[2] - left_x) ** 2 + (x[3] - top_y) ** 2)
-
-        fst_idx, fst_kind, left_x, top_y = candidates[0]
-        fst_bbox = subjects[fst_idx]['bbox'] if fst_kind == SUB_BIT_KIND else objects[fst_idx - OBJ_IDX_OFFSET]['bbox']
-        candidates.sort(
-            key=lambda x: bbox_distance(fst_bbox, subjects[x[0]]['bbox']) if x[1] == SUB_BIT_KIND else bbox_distance(
-                fst_bbox, objects[x[0] - OBJ_IDX_OFFSET]['bbox']))
-        nxt = None
-
-        for i in range(1, len(candidates)):
-            if candidates[i][1] ^ fst_kind == 1:
-                nxt = candidates[i]
-                break
-        if nxt is None:
-            break
-
-        if fst_kind == SUB_BIT_KIND:
-            sub_idx, obj_idx = fst_idx, nxt[0] - OBJ_IDX_OFFSET
-        else:
-            sub_idx, obj_idx = nxt[0], fst_idx - OBJ_IDX_OFFSET
-
-        pair_dis = bbox_distance(subjects[sub_idx]["bbox"], objects[obj_idx]["bbox"])
-        nearest_dis = float("inf")
-        for i in range(N):
-            # 取消原先算法中 1对1 匹配的偏置
-            # if i in seen_idx or i == sub_idx:continue
-            nearest_dis = min(nearest_dis, bbox_distance(subjects[i]["bbox"], objects[obj_idx]["bbox"]))
-
-        if pair_dis >= 3 * nearest_dis:
-            seen_idx.add(sub_idx)
-            continue
-
-        seen_idx.add(sub_idx)
-        seen_idx.add(obj_idx + OBJ_IDX_OFFSET)
-        seen_sub_idx.add(sub_idx)
-
-        ret.append(
-            {
-                "sub_bbox": extract_subject_func(subjects[sub_idx]),
-                "obj_bboxes": [extract_object_func(objects[obj_idx])],
-                "sub_idx": sub_idx,
-            }
-        )
-
-    for i in range(len(objects)):
-        j = i + OBJ_IDX_OFFSET
-        if j in seen_idx:
-            continue
-        seen_idx.add(j)
-        nearest_dis, nearest_sub_idx = float("inf"), -1
-        for k in range(len(subjects)):
-            dis = bbox_distance(objects[i]["bbox"], subjects[k]["bbox"])
-            if dis < nearest_dis:
-                nearest_dis = dis
-                nearest_sub_idx = k
-
-        for k in range(len(subjects)):
-            if k != nearest_sub_idx:
-                continue
-            if k in seen_sub_idx:
-                for kk in range(len(ret)):
-                    if ret[kk]["sub_idx"] == k:
-                        ret[kk]["obj_bboxes"].append(extract_object_func(objects[i]))
-                        break
-            else:
-                ret.append(
-                    {
-                        "sub_bbox": extract_subject_func(subjects[k]),
-                        "obj_bboxes": [extract_object_func(objects[i])],
-                        "sub_idx": k,
-                    }
-                )
-            seen_sub_idx.add(k)
-            seen_idx.add(k)
-
-    for i in range(len(subjects)):
-        if i in seen_sub_idx:
-            continue
-        ret.append(
-            {
-                "sub_bbox": extract_subject_func(subjects[i]),
-                "obj_bboxes": [],
-                "sub_idx": i,
-            }
-        )
-
-    return ret

mineru/version.py

@@ -1 +1 @@
-__version__ = "2.1.3"
+__version__ = "2.1.1"

signatures/version1/cla.json

@@ -399,6 +399,270 @@
       "created_at": "2025-07-22T03:04:40Z",
       "repoId": 765083837,
       "pullRequestNo": 3129
+    },
+    {
+      "name": "jinghuan-Chen",
+      "id": 42742857,
+      "comment_id": 3114162786,
+      "created_at": "2025-07-24T16:49:20Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3175
+    },
+    {
+      "name": "androllen",
+      "id": 5212108,
+      "comment_id": 3124534114,
+      "created_at": "2025-07-27T16:47:06Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3190
+    },
+    {
+      "name": "SirlyDreamer",
+      "id": 45280500,
+      "comment_id": 3138729334,
+      "created_at": "2025-07-31T06:36:17Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3222
+    },
+    {
+      "name": "yeahjack",
+      "id": 48432972,
+      "comment_id": 3155914249,
+      "created_at": "2025-08-05T16:59:05Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3269
+    },
+    {
+      "name": "loveRhythm1990",
+      "id": 52204121,
+      "comment_id": 3166629630,
+      "created_at": "2025-08-08T05:32:33Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3281
+    },
+    {
+      "name": "sleepyy-dog",
+      "id": 203856888,
+      "comment_id": 3209716785,
+      "created_at": "2025-08-21T09:23:23Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3354
+    },
+    {
+      "name": "147phoenix",
+      "id": 136807090,
+      "comment_id": 3294702328,
+      "created_at": "2025-09-16T03:18:30Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3477
+    },
+    {
+      "name": "e06084",
+      "id": 24450527,
+      "comment_id": 3302810047,
+      "created_at": "2025-09-17T12:42:23Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3489
+    },
+    {
+      "name": "cjsdurj",
+      "id": 29703648,
+      "comment_id": 3385821350,
+      "created_at": "2025-10-09T13:13:49Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3672
+    },
+    {
+      "name": "ye7love7",
+      "id": 72247118,
+      "comment_id": 3393172470,
+      "created_at": "2025-10-11T10:34:06Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3699
+    },
+    {
+      "name": "magicyuan876",
+      "id": 72277650,
+      "comment_id": 3408782163,
+      "created_at": "2025-10-16T01:09:35Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3739
+    },
+    {
+      "name": "yongtenglei",
+      "id": 65058560,
+      "comment_id": 3408954858,
+      "created_at": "2025-10-16T02:42:08Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3740
+    },
+    {
+      "name": "xvlincaigou",
+      "id": 123342936,
+      "comment_id": 3446786700,
+      "created_at": "2025-10-25T14:38:54Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3841
+    },
+    {
+      "name": "pzc163",
+      "id": 105222039,
+      "comment_id": 3449251055,
+      "created_at": "2025-10-27T02:17:34Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3842
+    },
+    {
+      "name": "aopstudio",
+      "id": 32569353,
+      "comment_id": 3460341998,
+      "created_at": "2025-10-29T08:27:03Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3870
+    },
+    {
+      "name": "Flynn-Zh",
+      "id": 23734538,
+      "comment_id": 3510150100,
+      "created_at": "2025-11-10T08:32:38Z",
+      "repoId": 765083837,
+      "pullRequestNo": 3966
+    },
+    {
+      "name": "eric-ozim",
+      "id": 233696103,
+      "comment_id": 3562698108,
+      "created_at": "2025-11-21T11:56:22Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4040
+    },
+    {
+      "name": "zyileven",
+      "id": 40888939,
+      "comment_id": 3579403940,
+      "created_at": "2025-11-26T06:12:20Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4066
+    },
+    {
+      "name": "borderlayout",
+      "id": 19464348,
+      "comment_id": 3626366462,
+      "created_at": "2025-12-08T11:07:40Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4141
+    },
+    {
+      "name": "lc345",
+      "id": 67670904,
+      "comment_id": 3659976666,
+      "created_at": "2025-12-16T11:03:21Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4196
+    },
+    {
+      "name": "RIORYO",
+      "id": 39397888,
+      "comment_id": 3701766690,
+      "created_at": "2025-12-31T08:54:17Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4277
+    },
+    {
+      "name": "kingdomad",
+      "id": 34766852,
+      "comment_id": 3707777776,
+      "created_at": "2026-01-04T06:16:13Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4283
+    },
+    {
+      "name": "tommygood",
+      "id": 96759292,
+      "comment_id": 3748312307,
+      "created_at": "2026-01-14T08:05:44Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4365
+    },
+    {
+      "name": "pgoslatara",
+      "id": 14027534,
+      "comment_id": 3799210889,
+      "created_at": "2026-01-26T11:58:59Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4421
+    },
+    {
+      "name": "guguducken",
+      "id": 22561920,
+      "comment_id": 3815970299,
+      "created_at": "2026-01-29T07:33:53Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4435
+    },
+    {
+      "name": "Arrmsgt",
+      "id": 115080429,
+      "comment_id": 3851141607,
+      "created_at": "2026-02-05T05:17:32Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4498
+    },
+    {
+      "name": "wzgrx",
+      "id": 39661556,
+      "comment_id": 3854380968,
+      "created_at": "2026-02-05T15:26:44Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4504
+    },
+    {
+      "name": "marswen",
+      "id": 24496561,
+      "comment_id": 3971065560,
+      "created_at": "2026-02-27T06:26:29Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4555
+    },
+    {
+      "name": "vanchy-z",
+      "id": 63965264,
+      "comment_id": 3976573836,
+      "created_at": "2026-02-28T07:14:07Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4560
+    },
+    {
+      "name": "troyhantech",
+      "id": 92877246,
+      "comment_id": 4088446017,
+      "created_at": "2026-03-19T07:58:56Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4631
+    },
+    {
+      "name": "vivekvar-dl",
+      "id": 69627205,
+      "comment_id": 4103182168,
+      "created_at": "2026-03-21T11:54:30Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4636
+    },
+    {
+      "name": "vivekvar-dl",
+      "id": 69627205,
+      "comment_id": 4103186013,
+      "created_at": "2026-03-21T11:57:40Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4636
+    },
+    {
+      "name": "UaRuairc",
+      "id": 77743840,
+      "comment_id": 4118931754,
+      "created_at": "2026-03-24T14:51:59Z",
+      "repoId": 765083837,
+      "pullRequestNo": 4654
     }
   ]
 }