[DBNET]-Bad performance on long text detection

[fatfishZhao]

Hi, thanks for your great job.
I'm using R50DCN dbnet for chinese text detection. I used about 10k pictures for training based on the pretrain model.
When testing, long text cannot be detected, some examples are in the bottom.
Can you give me some explanation of this performance? How can I fix this problem?

[0xCreo]

Same Issue. What's the problem of the origin config?

[fatfishZhao]

Solved. There are 2 problems.

The first one is the original shrink and unclip methods in paper, which is not suitable for long text ( the unclipped box is thinner than ground truth), so I changed these methods by my understanding.

The second one is a bug in the

mmocr/mmocr/models/textdet/postprocess/wrapper.py

Line 215 in 80741e1

epsilon = 0.01 * cv2.arcLength(poly, True)

0.01 is too big for long texts. The program will get two points rather than 4 points under this setting.
Set this value to 0.002 is much better.

[gaotongxiao]

Thanks for sharing your solution. Does the final performance look all good?

[fatfishZhao]

Yes, much better.

[fatfishZhao]

Actually, I'm confused about PPOCR's results. They also use dbnet and the presentation pictures in the repo is pretty good on long texts. The only different I found in the code between mmocr and ppocr is they use bigger unclip ratio.

[gaotongxiao]

It's probably because PPOCR uses much more private training data...

[Sanster]

I wrote a blog about this issue, if anyone is interested in this issue, check it out link

[fatfishZhao]

Here is my replacement for shrink and unclip method.
Using A and L to calculate the font size, and set the shrink/unclip distance to be a fixed ratio of the font size.

Note: r is same when shrinking and unclipping.

[0xCreo]

Here is my replacement for shrink and unclip method. Using A and L to calculate the font size, and set the shrink/unclip distance to be a fixed ratio of the font size. Note: r is same when shrinking and unclipping.

I write a simple test for your formula, but it doesn't work well in different h/w ratio. Did I make a mistake here?

[fatfishZhao]

I think so. It works well in my project.

[viviayi]

I think so. It works well in my project.

Hello, Thank you for sharing your method. I tried in my project, it works well for long text, but I found it unclip too much on short text. Do you have same problem? How you fixed it?

[fatfishZhao]

I think so. It works well in my project.

Hello, Thank you for sharing your method. I tried in my project, it works well for long text, but I found it unclip too much on short text. Do you have same problem? How you fixed it?

Hi, what is your r setting, you can try smaller r than 0.4 in paper, like 0.2.

[qiuzhixin9527]

@Sanster @fatfishZhao @viviayi @gaotongxiao #
Bad performance on long text detection,

it is my config
model = dict(
type='DBNet',
backbone=dict(
type='mmdet.ResNet',
depth=50,
num_stages=4,
out_indices=(0, 1, 2, 3),
frozen_stages=-1,
norm_cfg=dict(type='BN', requires_grad=True),
norm_eval=False,
style='pytorch',
dcn=dict(type='DCNv2', deform_groups=1, fallback_on_stride=False),
init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50'),
stage_with_dcn=(False, True, True, True)),
neck=dict(
type='FPNC',
in_channels=[256, 512, 1024, 2048],
lateral_channels=256,
asf_cfg=dict(attention_type='ScaleChannelSpatial')),
det_head=dict(
type='DBHead',
in_channels=256,
module_loss=dict(type='DBModuleLoss'),
postprocessor=dict(
type='DBPostprocessor', text_repr_type='quad',
epsilon_ratio=0.002)),
data_preprocessor=dict(
type='TextDetDataPreprocessor',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
bgr_to_rgb=True,
pad_size_divisor=32))
train_pipeline = [
dict(type='LoadImageFromFile', color_type='color_ignore_orientation'),
dict(
type='LoadOCRAnnotations',
with_bbox=True,
with_polygon=True,
with_label=True),
dict(
type='TorchVisionWrapper',
op='ColorJitter',
brightness=0.12549019607843137,
saturation=0.5),
dict(
type='ImgAugWrapper',
args=[['Fliplr', 0.5], {
'cls': 'Affine',
'rotate': [-10, 10]
}, ['Resize', [0.5, 3.0]]]),
dict(type='RandomCrop', min_side_ratio=0.1),
dict(type='Resize', scale=(1024, 1024), keep_ratio=True),
dict(type='Pad', size=(1024, 1024)),
dict(
type='PackTextDetInputs',
meta_keys=('img_path', 'ori_shape', 'img_shape'))
]
test_pipeline = [
dict(type='LoadImageFromFile', color_type='color_ignore_orientation'),
dict(type='Resize', scale=(1024, 1024), keep_ratio=True),
dict(
type='LoadOCRAnnotations',
with_polygon=True,
with_bbox=True,
with_label=True),
dict(
type='PackTextDetInputs',
meta_keys=('img_path', 'ori_shape', 'img_shape', 'scale_factor',
'instances'))
]
default_scope = 'mmocr'
env_cfg = dict(
cudnn_benchmark=False,
mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
dist_cfg=dict(backend='nccl'))
randomness = dict(seed=None)
default_hooks = dict(
timer=dict(type='IterTimerHook'),
logger=dict(type='LoggerHook', interval=5),
param_scheduler=dict(type='ParamSchedulerHook'),
checkpoint=dict(type='CheckpointHook', interval=20),
sampler_seed=dict(type='DistSamplerSeedHook'),
sync_buffer=dict(type='SyncBuffersHook'),
visualization=dict(
type='VisualizationHook',
interval=1,
enable=False,
show=False,
draw_gt=False,
draw_pred=False))
log_level = 'INFO'
log_processor = dict(type='LogProcessor', window_size=10, by_epoch=True)
load_from = 'https://download.openmmlab.com/mmocr/textdet/dbnetpp/tmp_1.0_pretrain/dbnetpp_r50dcnv2_fpnc_100k_iter_synthtext-20220502-352fec8a.pth'
resume = False
val_evaluator = dict(type='HmeanIOUMetric')
test_evaluator = dict(type='HmeanIOUMetric')
vis_backends = [dict(type='LocalVisBackend')]
visualizer = dict(
type='TextDetLocalVisualizer',
name='visualizer',
vis_backends=[dict(type='LocalVisBackend')])
icdar2015_textdet_data_root = '/home/aipf/work/团队共享目录/zhixin/datasets/icdar2015'
icdar2015_textdet_train = dict(
type='OCRDataset',
data_root='/home/aipf/work/团队共享目录/zhixin/datasets/icdar2015',
ann_file='textdet_train.json',
filter_cfg=dict(filter_empty_gt=True, min_size=32),
pipeline=None)
icdar2015_textdet_test = dict(
type='OCRDataset',
data_root='/home/aipf/work/团队共享目录/zhixin/datasets/icdar2015',
ann_file='textdet_test.json',
test_mode=True,
pipeline=None)
optim_wrapper = dict(
type='OptimWrapper',
optimizer=dict(type='SGD', lr=0.0035, momentum=0.9, weight_decay=0.0001))
train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=1200, val_interval=20)
val_cfg = dict(type='ValLoop')
test_cfg = dict(type='TestLoop')
param_scheduler = [dict(type='PolyLR', power=0.9, eta_min=1e-07, end=200)]
train_list = [
dict(
type='OCRDataset',
data_root='/home/aipf/work/团队共享目录/zhixin/datasets/icdar2015',
ann_file='textdet_train.json',
filter_cfg=dict(filter_empty_gt=True, min_size=32),
pipeline=None)
]
test_list = [
dict(
type='OCRDataset',
data_root='/home/aipf/work/团队共享目录/zhixin/datasets/icdar2015',
ann_file='textdet_test.json',
test_mode=True,
pipeline=None)
]
train_dataloader = dict(
batch_size=8,
num_workers=8,
persistent_workers=True,
sampler=dict(type='DefaultSampler', shuffle=True),
dataset=dict(
type='ConcatDataset',
datasets=[
dict(
type='OCRDataset',
data_root='/home/aipf/work/团队共享目录/zhixin/datasets/icdar2015',
ann_file='textdet_train.json',
filter_cfg=dict(filter_empty_gt=True, min_size=32),
pipeline=None)
],
pipeline=[
dict(
type='LoadImageFromFile',
color_type='color_ignore_orientation'),
dict(
type='LoadOCRAnnotations',
with_bbox=True,
with_polygon=True,
with_label=True),
dict(
type='TorchVisionWrapper',
op='ColorJitter',
brightness=0.12549019607843137,
saturation=0.5),
dict(
type='ImgAugWrapper',
args=[['Fliplr', 0.5], {
'cls': 'Affine',
'rotate': [-10, 10]
}, ['Resize', [0.5, 3.0]]]),
dict(type='RandomCrop', min_side_ratio=0.1),
dict(type='Resize', scale=(1024, 1024), keep_ratio=True),
dict(type='Pad', size=(1024, 1024)),
dict(
type='PackTextDetInputs',
meta_keys=('img_path', 'ori_shape', 'img_shape'))
]))
val_dataloader = dict(
batch_size=8,
num_workers=8,
persistent_workers=True,
sampler=dict(type='DefaultSampler', shuffle=False),
dataset=dict(
type='ConcatDataset',
datasets=[
dict(
type='OCRDataset',
data_root='/home/aipf/work/团队共享目录/zhixin/datasets/icdar2015',
ann_file='textdet_test.json',
test_mode=True,
pipeline=None)
],
pipeline=[
dict(
type='LoadImageFromFile',
color_type='color_ignore_orientation'),
dict(type='Resize', scale=(1024, 1024), keep_ratio=True),
dict(
type='LoadOCRAnnotations',
with_polygon=True,
with_bbox=True,
with_label=True),
dict(
type='PackTextDetInputs',
meta_keys=('img_path', 'ori_shape', 'img_shape',
'scale_factor', 'instances'))
]))
test_dataloader = dict(
batch_size=8,
num_workers=8,
persistent_workers=True,
sampler=dict(type='DefaultSampler', shuffle=False),
dataset=dict(
type='ConcatDataset',
datasets=[
dict(
type='OCRDataset',
data_root='/home/aipf/work/团队共享目录/zhixin/datasets/icdar2015',
ann_file='textdet_test.json',
test_mode=True,
pipeline=None)
],
pipeline=[
dict(
type='LoadImageFromFile',
color_type='color_ignore_orientation'),
dict(type='Resize', scale=(1024, 1024), keep_ratio=True),
dict(
type='LoadOCRAnnotations',
with_polygon=True,
with_bbox=True,
with_label=True),
dict(
type='PackTextDetInputs',
meta_keys=('img_path', 'ori_shape', 'img_shape',
'scale_factor', 'instances'))
]))
launcher = 'none'
work_dir = 'output/dbpp0417_dcnv2'