[Dataset Preparer] MJSynth (#1791)

* finialize

* finialize

configs/textrecog/_base_/datasets/mjsynth.py

@@ -1,17 +1,7 @@
-mjsynth_textrecog_data_root = 'data/rec/Syn90k/'
+mjsynth_textrecog_data_root = 'data/mjsynth'
 
-mjsynth_textrecog_test = dict(
+mjsynth_textrecog_train = dict(
     type='OCRDataset',
     data_root=mjsynth_textrecog_data_root,
-    data_prefix=dict(img_path='mnt/ramdisk/max/90kDICT32px'),
+    ann_file='textrecog_train.json',
-    ann_file='train_labels.json',
-    test_mode=False,
-    pipeline=None)
-
-mjsynth_sub_textrecog_train = dict(
-    type='OCRDataset',
-    data_root=mjsynth_textrecog_data_root,
-    data_prefix=dict(img_path='mnt/ramdisk/max/90kDICT32px'),
-    ann_file='subset_train_labels.json',
-    test_mode=False,
     pipeline=None)

configs/textrecog/abinet/abinet-vision_20e_st-an_mj.py

@@ -24,7 +24,7 @@ param_scheduler = [
 
 # dataset settings
 train_list = [
-    _base_.mjsynth_textrecog_test, _base_.synthtext_an_textrecog_train
+    _base_.mjsynth_textrecog_train, _base_.synthtext_an_textrecog_train
 ]
 test_list = [
     _base_.cute80_textrecog_test, _base_.iiit5k_textrecog_test,

configs/textrecog/abinet/abinet_20e_st-an_mj.py

@@ -26,7 +26,7 @@ param_scheduler = [
 
 # dataset settings
 train_list = [
-    _base_.mjsynth_textrecog_test, _base_.synthtext_an_textrecog_train
+    _base_.mjsynth_textrecog_train, _base_.synthtext_an_textrecog_train
 ]
 test_list = [
     _base_.cute80_textrecog_test, _base_.iiit5k_textrecog_test,

configs/textrecog/aster/aster_resnet45_6e_st_mj.py

@@ -15,7 +15,7 @@ _base_ = [
 
 # dataset settings
 train_list = [
-    _base_.mjsynth_textrecog_test,
+    _base_.mjsynth_textrecog_train,
     _base_.synthtext_textrecog_train,
 ]
 test_list = [

configs/textrecog/crnn/crnn_mini-vgg_5e_mj.py

@@ -12,7 +12,7 @@ _base_ = [
     '_base_crnn_mini-vgg.py',
 ]
 # dataset settings
-train_list = [_base_.mjsynth_textrecog_test]
+train_list = [_base_.mjsynth_textrecog_train]
 test_list = [
     _base_.cute80_textrecog_test, _base_.iiit5k_textrecog_test,
     _base_.svt_textrecog_test, _base_.svtp_textrecog_test,

configs/textrecog/master/master_resnet31_12e_st_mj_sa.py

@@ -23,7 +23,7 @@ param_scheduler = [
 
 # dataset settings
 train_list = [
-    _base_.mjsynth_textrecog_test, _base_.synthtext_textrecog_train,
+    _base_.mjsynth_textrecog_train, _base_.synthtext_textrecog_train,
     _base_.synthtext_add_textrecog_train
 ]
 test_list = [

configs/textrecog/nrtr/nrtr_modality-transform_6e_st_mj.py

@@ -20,7 +20,7 @@ param_scheduler = [
 ]
 
 # dataset settings
-train_list = [_base_.mjsynth_textrecog_test, _base_.synthtext_textrecog_train]
+train_list = [_base_.mjsynth_textrecog_train, _base_.synthtext_textrecog_train]
 test_list = [
     _base_.cute80_textrecog_test, _base_.iiit5k_textrecog_test,
     _base_.svt_textrecog_test, _base_.svtp_textrecog_test,

configs/textrecog/nrtr/nrtr_resnet31-1by16-1by8_6e_st_mj.py

@@ -20,7 +20,7 @@ param_scheduler = [
 ]
 
 # dataset settings
-train_list = [_base_.mjsynth_textrecog_test, _base_.synthtext_textrecog_train]
+train_list = [_base_.mjsynth_textrecog_train, _base_.synthtext_textrecog_train]
 test_list = [
     _base_.cute80_textrecog_test, _base_.iiit5k_textrecog_test,
     _base_.svt_textrecog_test, _base_.svtp_textrecog_test,

configs/textrecog/satrn/satrn_shallow_5e_st_mj.py

@@ -13,7 +13,7 @@ _base_ = [
 ]
 
 # dataset settings
-train_list = [_base_.mjsynth_textrecog_test, _base_.synthtext_textrecog_train]
+train_list = [_base_.mjsynth_textrecog_train, _base_.synthtext_textrecog_train]
 test_list = [
     _base_.cute80_textrecog_test, _base_.iiit5k_textrecog_test,
     _base_.svt_textrecog_test, _base_.svtp_textrecog_test,

configs/textrecog/svtr/svtr-tiny_20e_st_mj.py

@@ -41,7 +41,7 @@ param_scheduler = [
 ]
 
 # dataset settings
-train_list = [_base_.mjsynth_textrecog_test, _base_.synthtext_textrecog_train]
+train_list = [_base_.mjsynth_textrecog_train, _base_.synthtext_textrecog_train]
 test_list = [
     _base_.cute80_textrecog_test, _base_.iiit5k_textrecog_test,
     _base_.svt_textrecog_test, _base_.svtp_textrecog_test,

dataset_zoo/mjsynth/metafile.yml

@@ -0,0 +1,37 @@
+Name: 'Synthetic Word Dataset (MJSynth/Syn90k)'
+Paper:
+  Title: Reading Text in the Wild with Convolutional Neural Networks
+  URL: https://arxiv.org/pdf/1412.1842.pdf
+  Venue: International Journal of Computer Vision
+  Year: '2016'
+  BibTeX: '@InProceedings{Jaderberg14c,
+  author       = "Max Jaderberg and Karen Simonyan and Andrea Vedaldi and Andrew Zisserman",
+  title        = "Synthetic Data and Artificial Neural Networks for Natural Scene Text Recognition",
+  booktitle    = "Workshop on Deep Learning, NIPS",
+  year         = "2014",
+  }
+
+  @Article{Jaderberg16,
+  author       = "Max Jaderberg and Karen Simonyan and Andrea Vedaldi and Andrew Zisserman",
+  title        = "Reading Text in the Wild with Convolutional Neural Networks",
+  journal      = "International Journal of Computer Vision",
+  number       = "1",
+  volume       = "116",
+  pages        = "1--20",
+  month        = "jan",
+  year         = "2016",
+  }'
+Data:
+  Website: https://www.robots.ox.ac.uk/~vgg/data/text/
+  Language:
+    - English
+  Scene:
+    - Synthesis
+  Granularity:
+    - Word
+  Tasks:
+    - textrecog
+  License:
+    Type: N/A
+    Link: N/A
+  Format: .txt

dataset_zoo/mjsynth/sample_anno.md

@@ -0,0 +1,6 @@
+**Text Recognition**
+
+```txt
+./3000/7/182_slinking_71711.jpg 71711
+./3000/7/182_REMODELERS_64541.jpg 64541
+```

dataset_zoo/mjsynth/textrecog.py

@@ -0,0 +1,40 @@
+data_root = 'data/mjsynth'
+cache_path = 'data/cache'
+
+train_preparer = dict(
+    obtainer=dict(
+        type='NaiveDataObtainer',
+        cache_path=cache_path,
+        files=[
+            dict(
+                url='https://thor.robots.ox.ac.uk/~vgg/data/text/'
+                'mjsynth.tar.gz',
+                save_name='mjsynth.tar.gz',
+                md5='7bf2b60ad935eaf64e5b606f782d68e5',
+                split=['train'],
+                content=['image', 'annotation'],
+                mapping=[
+                    [
+                        'mjsynth/mnt/ramdisk/max/90kDICT32px/*/',
+                        'textrecog_imgs/train/'
+                    ],
+                    [
+                        'mjsynth/mnt/ramdisk/max/90kDICT32px/annotation.txt',
+                        'annotations/annotation.txt'
+                    ]
+                ]),
+        ]),
+    gatherer=dict(type='MonoGatherer', ann_name='annotation.txt'),
+    parser=dict(
+        type='ICDARTxtTextRecogAnnParser',
+        separator=' ',
+        format='img text',
+        remove_strs=None),
+    packer=dict(type='TextRecogPacker'),
+    dumper=dict(type='JsonDumper'),
+)
+
+delete = ['mjsynth', 'annotations']
+
+config_generator = dict(
+    type='TextRecogConfigGenerator', data_root=data_root, test_anns=None)

docs/en/user_guides/config.md

@@ -291,7 +291,7 @@ For example, for text recognition tasks, Syn90k is used as the training set, whi
 
 ```Python
 # text recognition dataset configuration
-mjsynth_textrecog_test = dict(
+mjsynth_textrecog_train = dict(
     type='OCRDataset',
     data_root='data/rec/Syn90k/',
     data_prefix=dict(img_path='mnt/ramdisk/max/90kDICT32px'),
@@ -374,7 +374,7 @@ train_dataloader = dict(
     sampler=dict(type='DefaultSampler', shuffle=True),
     dataset=dict(
         type='ConcatDataset',
-        datasets=[mjsynth_textrecog_test],
+        datasets=[mjsynth_textrecog_train],
         pipeline=train_pipeline))
 val_dataloader = dict(
     batch_size=1,

docs/en/user_guides/dataset_prepare.md

@@ -171,7 +171,7 @@ _base_ = [ # Import all dataset configurations you want to use
 ]
 
 # List of training datasets
-train_list = [_base_.mjsynth_textrecog_test]
+train_list = [_base_.mjsynth_textrecog_train]
 # List of testing datasets
 test_list = [
     _base_.cute80_textrecog_test, _base_.iiit5k_textrecog_test, _base_.svt_textrecog_test,

docs/zh_cn/user_guides/config.md

@@ -288,7 +288,7 @@ test_cfg = dict(type='TestLoop')
 
 ```Python
 # 识别数据集配置
-mjsynth_textrecog_test = dict(
+mjsynth_textrecog_train = dict(
     type='OCRDataset',
     data_root='data/rec/Syn90k/',
     data_prefix=dict(img_path='mnt/ramdisk/max/90kDICT32px'),
@@ -373,7 +373,7 @@ train_dataloader = dict(
     sampler=dict(type='DefaultSampler', shuffle=True),
     dataset=dict(
         type='ConcatDataset',
-        datasets=[mjsynth_textrecog_test],
+        datasets=[mjsynth_textrecog_train],
         pipeline=train_pipeline))
 val_dataloader = dict(
     batch_size=1,

docs/zh_cn/user_guides/dataset_prepare.md

@@ -171,7 +171,7 @@ _base_ = [ # 导入所有需要使用的数据集配置
 ]
 
 # 训练集列表
-train_list = [_base_.mjsynth_textrecog_test]
+train_list = [_base_.mjsynth_textrecog_train]
 # 测试集列表
 test_list = [
     _base_.cute80_textrecog_test, _base_.iiit5k_textrecog_test, _base_.svt_textrecog_test,

mmocr/datasets/preparers/obtainers/naive_data_obtainer.py

@@ -186,7 +186,8 @@ class NaiveDataObtainer:
             if '*' in src:
                 mkdir_or_exist(dst)
                 for f in glob.glob(src):
-                    if not osp.exists(osp.join(dst, osp.basename(f))):
+                    if not osp.exists(
+                            osp.join(dst, osp.relpath(f, self.data_root))):
                         shutil.move(f, dst)
 
             elif osp.exists(src) and not osp.exists(dst):

mmocr/datasets/preparers/packers/textrecog_packer.py

@@ -50,11 +50,8 @@ class TextRecogPacker(BasePacker):
         """
 
         img_name, text = sample
-        # TODO: remove hard code
+        img_name = osp.relpath(img_name, self.data_root)
-        packed_instance = dict(
+        packed_instance = dict(instances=[dict(text=text)], img_path=img_name)
-            instances=[dict(text=text)],
-            img_path=osp.join('textrecog_imgs', self.split,
-                              osp.basename(img_name)))
 
         return packed_instance
 

mmocr/datasets/preparers/parsers/icdar_txt_parser.py

@@ -122,6 +122,6 @@ class ICDARTxtTextRecogAnnParser(BaseParser):
             if text == self.ignore:
                 continue
             img_name = anno['img']
-            samples.append((osp.join(img_dir, osp.basename(img_name)), text))
+            samples.append((osp.join(img_dir, img_name), text))
 
         return samples

tests/test_datasets/test_preparers/test_packers/test_textrecog_packer.py

@@ -13,14 +13,11 @@ class TestTextRecogPacker(unittest.TestCase):
 
     def test_pack_instance(self):
 
-        packer = TextRecogPacker(data_root='', split='test')
+        packer = TextRecogPacker(data_root='data/test/', split='test')
-        sample = ('test.jpg', 'text')
+        sample = ('data/test/test.jpg', 'text')
         results = packer.pack_instance(sample)
         self.assertDictEqual(
-            results,
+            results, dict(img_path='test.jpg', instances=[dict(text='text')]))
-            dict(
-                img_path=osp.join('textrecog_imgs', 'test', 'test.jpg'),
-                instances=[dict(text='text')]))
 
     def test_add_meta(self):
         packer = TextRecogPacker(data_root='', split='test')