弯曲文字检测报错The requested array has an inhomogeneous shape after 1 dimensions

[blvyoucan]

请提供下述完整信息以便快速定位问题/Please provide the following information to quickly locate the problem

• 系统环境/System Environment：ubuntu 20.04
• 版本号/Version：Paddle： 2.6 PaddleOCR：dygraph 问题相关组件/Related components：ppocr/postprocess/db_postprocess.py
• 运行指令/Command Code：python tools/train.py -c configs/det/det_r50_db++_icdar15.yml
• 完整报错/Complete Error Message：eval model:: 1%|▍ | 11/2000 [00:01<03:01, 10.95it/s]Traceback (most recent call last): File "/home/suma/llm/train/PaddleOCR/tools/train.py", line 227, in main(config, device, logger, vdl_writer, seed) File "/home/suma/llm/train/PaddleOCR/tools/train.py", line 198, in main program.train(config, train_dataloader, valid_dataloader, device, model, File "/home/suma/llm/train/PaddleOCR/tools/program.py", line 392, in train cur_metric = eval( File "/home/suma/llm/train/PaddleOCR/tools/program.py", line 570, in eval post_result = post_process_class(preds, batch_numpy[1]) File "/home/suma/llm/train/PaddleOCR/ppocr/postprocess/db_postprocess.py", line 237, in call boxes, scores = self.polygons_from_bitmap(pred[batch_index], File "/home/suma/llm/train/PaddleOCR/ppocr/postprocess/db_postprocess.py", line 84, in polygons_from_bitmap box = self.unclip(points, self.unclip_ratio) File "/home/suma/llm/train/PaddleOCR/ppocr/postprocess/db_postprocess.py", line 156, in unclip expanded = np.array(offset.Execute(distance)) ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (2,) + inhomogeneous part.

将det_r50_db++_icdar15.yml文件的

# 第57行
det_box_type: 'quad'
# 改为了
box_type: 'poly'

问题同 #11369

调试得到的如下数据

box = array([[ 434,  903],
       [ 507,  942],
       [ 698,  999],
       [ 984, 1037],
       [1206, 1019],
       [1344,  981],
       [ 971, 1012],
       [ 688,  971],
       [ 647,  948],
       [ 652,  938],
       [1338,  956],
       [1362,  979],
       [1419,  959],
       [1625,  952],
       [1641,  937],
       [1628,  921],
       [1387,  936],
       [ 582,  921],
       [ 478,  887],
       [ 444,  889]], dtype=int32)

unclip_radio = 1.5

distance = 18.020694170941756

offset.Execute(distance) = [
[[484, 870], [585, 903], [1386, 918], [1627, 903], [1633, 904], [1638, 906], [1642, 910], [1655, 926], [1658, 931], [1659, 937], [1658, 943], [1653, 950], [1637, 965], [1632, 968], [1626, 970], [1422, 977], [1368, 996], [1362, 997], [1356, 996], [1355, 995], [1354, 996], [1349, 998], [1211, 1036], [1207, 1037], [985, 1055], [982, 1055], [696, 1017], [693, 1016], [502, 959], [499, 958], [426, 919], [421, 915], [417, 910], [416, 904], [417, 898], [419, 893], [429, 879], [434, 874], [439, 872], [443, 871], [477, 869]],

[[973, 994], [1241, 971], [725, 958]]
]

尝试了以下数据集

https://aistudio.baidu.com/datasetdetail/199400

https://aistudio.baidu.com/datasetdetail/208469

[cuicheng01]

默认的示例数据集可以吗

[blvyoucan]

默认的示例数据集可以吗

好像只提供了标注方法，没有提供示例数据集

[anthonyAndchen]

可以调整数据传递给 NumPy 数组的所有子列表，使其长度相同，以确保一致性。在ppocr/postprocess/db_postprocess.py的156行调整一下代码：

        data_array = offset.Execute(distance)
        max_length = max(len(sublist) for sublist in data_array)
        for sublist in data_array:
            while len(sublist) < max_length:
                sublist.append(sublist[-1])
        expanded = np.array(data_array)

测试后不再出现上述问题。

[blvyoucan]

可以调整数据传递给 NumPy 数组的所有子列表，使其长度相同，以确保一致性。在ppocr/postprocess/db_postprocess.py的156行调整一下代码：

        data_array = offset.Execute(distance)
        max_length = max(len(sublist) for sublist in data_array)
        for sublist in data_array:
            while len(sublist) < max_length:
                sublist.append(sublist[-1])
        expanded = np.array(data_array)

测试后不再出现上述问题。

报错是没有了，但是训练时评估的精度一直下降，不会上升

[UserWangZz]

def unclip(self, box, unclip_ratio):
        try:
            poly = Polygon(box)
            distance = poly.area * unclip_ratio / poly.length
            offset = pyclipper.PyclipperOffset()
            offset.AddPath(box, pyclipper.JT_ROUND, pyclipper.ET_CLOSEDPOLYGON)
            offset_polygon = offset.Execute(distance)
            expanded = np.array(offset_polygon)
        except Exception as e:
            print("error:",e)
            print("box:",box)
            print("distance:",distance)
            print("offset_polygon:",offset_polygon)
            return np.array(offset_polygon[0])
        return expanded

目前建议先改为这样，这是因为db对文本分割区域的异常导致的问题，在执行offset.Execute(distance)后，可能会返回两个长度不一致的numpy数组，导致np.array()报错。 下面是我可视化出来的一个异常box形状：