Create Davar-OCR

README.md

@@ -1 +1,57 @@
-# Davar-Lab OCR
+# DAVAR-OCR
+
+This the OCR repository of [DAVAR Lab](https://davar-lab.github.io/), from Hikvision Research Institute, China. 
+
+We begin to maintain this code repository to release the official implementations of our recent academic publishments in OCR.
+
+> Note: Due to the policy limits of the company. All of the codes were re-implemented based on the open-source frameworks, [mmdetection-1.2.0](https://github.com/open-mmlab/mmdetection/releases/tag/v1.2.0) and [mmcv-0.4.2](https://github.com/open-mmlab/mmcv/releases/tag/v0.4.2), from [open-mmlab](https://github.com/open-mmlab "open-mmlab"). Therefore, the demonstrated results might be slightly different from the reported performances.
+
+## Implementations
+To date, we have released / will release the following algorithms:
+
+- MANGO (to be released) (AAAI 2021)
+
+- SPIN (to be released) (AAAI 2021)
+
+- FREE (to be released) (TIP 2020)
+
+- TRIE (to be released) (ACM MM 2020)
+
+- [Text Perceptron](third_party/text_perceptron/) (AAAI 2020)
+
+- YORO (to be released) (ACM MM 2019)
+
+## Development Environment
+
+| Basic Env   | version |
+| :---------- | ------- |
+| Python      | 3.6     |
+| cuda        | 10.0    |
+| cudnn       | 7.6.3   |
+| opencv      | 3.4.9   |
+| pytorch     | 1.3.0   |
+| torchvision | 0.4.1   |
+
+## Installation and Development Instruction 
+We keep the main part of mmdetection and mmcv exactly same with the official version. Each algorithm is stored under `mmdetection/third_party` in a separate directory structure.
+
+To Download the repository and install the mmdetection and mmcv, please follow the instructions:
+``` basic
+>>> git clone https://github.com/DAVAR-Lab/DAVAR-OCR.git
+>>> cd DAVAR-OCR/
+>>> bash setup.sh
+```
+If you want to run some model, you can uncomment the corresponding importing statement in `mmdetection/third_party/__init__.py` directly in develop mode.
+
+For example, if you want to use the model of Text Perceptron, you could uncomment the line of `from .text_perceptron import *` in `mmdetection/third_party/__init__.py`.
+
+Going to the specifc algorithm's directory to see more details.
+## License
+This project is released under the [Apache 2.0 license](mmdetection/third_party/LICENSE)
+
+## Copyright
+
+The copyright of corresponding contributions of our third-party modules belongs to *Davar-Lab, Hikvision Research Institute, China*, and other codes from open source repository follows the original distributive licenses.
+
+## Welcome to DAVAR-LAB!
+See [latest news](https://davar-lab.github.io/news/) in DAVAR-Lab.

demo/text_perceptron_det/config/tp_r50_3stages_enlarge.py

@@ -0,0 +1,183 @@
+"""
+#########################################################################
+# Copyright (c) Davar Lab @ Hikvision Research Institute. All rights reserved.
+# Filename       :    tp_r50_3stages_enlarge.py
+# Abstract       :    Model settings
+
+# Current Version:    1.0.0
+# Author         :    Liang Qiao
+# Date           :    2020-05-31
+#########################################################################
+"""
+import os
+
+# Absolute path to the 'third_party' directory
+third_party_path = '/path/to/mmdetection/third_party/'
+
+model = dict(
+    type='TextPerceptronDet',
+    # Pre-trained model, can be downloaded in the model zoo of mmdetection
+    pretrained='/path/to/pretrained_model/resnet50-19c8e357.pth',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=3,
+        out_indices=(0, 1, 2),
+        dilations=(1, 2, 1),   # enlarge receptive field in 8x feature map
+        strides=(1, 1, 2),
+        frozen_stages=-1,
+        style='pytorch'),
+    neck=dict(
+        type='FPN',
+        in_channels=[256, 512, 1024],
+        out_channels=256,
+        start_level=0,
+        add_extra_convs=True,
+        num_outs=3),
+    mask_head=dict(
+        type='TPHead',
+        in_channels=256,
+        conv_out_channels=256,
+        conv_cfg=None,
+        norm_cfg=None,
+        # All of the segmentation losses, including center text/ head/ tail/ top&bottom boundary
+        loss_seg=dict(type='DiceLoss', loss_weight=1.0),
+        # Corner regression in head region
+        loss_reg_head=dict(type='SmoothL1Loss', beta=0.11, loss_weight=0.1,
+                           reduction='sum'),
+        # Corner regression in tail region
+        loss_reg_tail=dict(type='SmoothL1Loss', beta=0.11, loss_weight=0.1,
+                           reduction='sum'),
+        # boundary offset regression in center text region
+        loss_reg_bond=dict(type='SmoothL1Loss', beta=0.11, loss_weight=0.01,
+                           reduction='sum'),
+        ),
+    shape_transform_module=dict(
+        type='PointsGeneration',
+        # Re-implenmented in C++ (You can implement it in CUDA for further speed up)
+        libname='tp_points_generate.so',
+        libdir=os.path.join(third_party_path,
+                            'text_perceptron/mmdet/models/shape_transform_module/lib/'),
+        # Parameters for points generating
+        filter_ratio=0.6,
+        thres_text=0.35,
+        thres_head=0.45,
+        thres_bond=0.35,
+        point_num=14
+    )
+)
+# training and testing settings
+train_cfg = dict()
+test_cfg = dict()
+
+dataset_type = 'CustomDataset'
+data_root = ''
+
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', with_bbox=True),
+    # Implemented in third_party/text_perceptron
+    dict(type='RandomRotate', angles=(-15, 15)),
+    dict(type='DavarResize', img_scale=[(640, 960), (1400, 1600)],
+         multiscale_mode='range', keep_ratio=True),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=32),
+    # Ground truth generation
+    dict(type='TPDataGeneration',
+        shrink_head_ratio=0.35,
+        shrink_bond_ratio=0.15,
+        ignore_ratio=0.6,
+        lib_name='tp_data.so',
+        lib_dir=os.path.join(third_party_path,
+                             'text_perceptron/mmdet/datasets/pipelines/lib/')),
+    dict(type='TPFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_masks'],
+         meta_keys=('filename', 'scale_factor', 'pad_shape')),
+]
+test_pipeline = [
+    dict(type='LoadTestImg'),
+    dict(
+        type='MultiScaleFlipAug',
+        #img_scale=(1250, 950),                # Testing scale for SCUT-CTW1500
+        img_scale=(1350, 950),                # Testing scale for Total-Text
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='DefaultFormatBundle'),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    imgs_per_gpu=2,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        ann_file=[
+                '/path/to/datalist/total_text_train_datalist.json'
+                #'/path/to/datalist/ctw1500_train_datalist.json'
+                  ],
+        img_prefix=[
+                '/path/to/Images/Total-Text/'
+                #'/path/to/Images/SCUT-CTW1500/'
+                ],
+        pipeline=train_pipeline
+        ),
+    # Not used
+    val=dict(
+        type=dataset_type,
+        ann_file=data_root + 'annotations/instances_val2017.json',
+        img_prefix=data_root + 'val2017/',
+        img_scale=(1333, 800),
+        img_norm_cfg=img_norm_cfg,
+        size_divisor=32,
+        flip_ratio=0,
+        with_mask=False,
+        with_crowd=False,
+        with_label=True),
+    test=dict(
+        type=dataset_type,
+        imgs_per_gpu=1,
+        ann_file=None,
+        img_prefix=None,
+        pipeline=test_pipeline
+        ))
+# optimizer
+optimizer = dict(type='SGD', lr=0.005, momentum=0.9, weight_decay=0.00005)
+optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
+
+# learning policy
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=2000,
+    warmup_ratio=1.0 / 3,
+    step=[8, 16, 24])
+
+# checkpoint saved path
+checkpoint_config = dict(interval=5,  filename_tmpl='checkpoint/tp_r50_tt_epoch_{}.pth')
+# yapf:disable
+log_config = dict(
+    interval=10,
+    hooks=[
+        dict(type='TextLoggerHook'),
+    ])
+# yapf:enable
+# runtime settings
+total_epochs = 30
+device_ids = range(8)
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+
+# Work space to save training log
+work_dir = 'path/to/save/log/'
+
+# 'Pretrained model on Synthtext'
+# you can simply load_from = 'path/to/tp_det_r50_tt_e25-45b1f5cf.pth' to fine-tune current model into a new domain
+load_from = None
+
+resume_from = None
+workflow = [('train', 1)]