text_spotting

Text Spotting

This repository contains inference and training code for Text Spotting models based on Mask R-CNN like networks. Models code is designed to enable ONNX* export (with custom operations) and inference on CPU via OpenVINO™.

Setup

Prerequisites

• Ubuntu* 16.04
• GCC* 7.4.0
• Python* 3.6 or newer
• PyTorch* (custom, see Installation section)
• CUDA* 10.1
• OpenVINO™ 2020.1 with Python API

Installation

Create and activate virtual environment:

bash init_venv.sh

NOTE on this step will be install custom version of Pytorch and torchvison from specific commit. For more information, see the init_venv.sh

Download Datasets

To be able to train networks and/or get quality metrics for pre-trained ones, one have to download one dataset at least.

• https://rrc.cvc.uab.es/ - ICDAR2013 (Focused Scene Text), ICDAR2013 (Incidental Scene Text), ICDAR2017 (MLT), ... .
• http://www.iapr-tc11.org/mediawiki/index.php/MSRA_Text_Detection_500_Database_(MSRA-TD500) MSRA-TD500.
• https://bgshih.github.io/cocotext/ COCO-Text.

Convert Datasets

Extract downloaded datasets in following images_folder: $WORK_DIR/openvino_training_extensions/pytorch_toolkit/text_spotting/data/coco

Convert extracted datasets to format that is used internally.

python3 tools/create_dataset.py --config datasets/dataset_train.json --output data/coco/IC13TRAINTEST_IC15TRAIN_MSRATD500TRAINTEST_COCOTEXTTRAINVAL.json
python3 tools/create_dataset.py --config datasets/dataset_test.json --output data/coco/IC15TEST.json

Examples of dataset_configuration.json can be found in openvino_training_extensions/pytorch_toolkit/text_spotting/datasets.

The structure of the folder with datasets:

texxt_spotting/data/coco/
    ├── coco-text
    ├── icdar2013
    ├── icdar2015
    ├── MSRA-TD500
    ├── IC13TRAINTEST_IC15TRAIN_MSRATD500TRAINTEST_COCOTEXTTRAINVAL.json
    └── IC15TEST.json

Training

To train Text Spotter model run:

python3 tools/train.py configs/text-spotting-0001.json

One can point to pre-trained model checkpoint inside configuration file to start training from pre-trained weights. Change configs/text-spotting-0001.json:

...
"checkpoint": "<path_to_weights>",
...

Known issue: 'Nan' in log output.

metrics/detection/cls_accuracy: 0.95204, metrics/rpn/cls_accuracy/0: 0.969265, metrics/rpn/cls_accuracy/1: 1.0,
metrics/rpn/cls_accuracy/2: 1.0, metrics/rpn/cls_accuracy/3: nan, metrics/rpn/cls_accuracy/4: nan, metrics/rpn/cls_precision/0: nan,
metrics/rpn/cls_precision/1: nan, metrics/rpn/cls_precision/2: nan, metrics/rpn/cls_precision/3: nan, metrics/rpn/cls_precision/4: nan,
metrics/rpn/cls_recall/0: nan, metrics/rpn/cls_recall/1: nan, metrics/rpn/cls_recall/2: nan, metrics/rpn/cls_recall/3: nan,
time elapsed/~left: 0:34:33 / 2 days, 7:37:14 (1.01 sec/it)
WARNING 17-01-20 13:42:25 x2num.py:  14] NaN or Inf found in input tensor.

Evaluation

tools/test.py script is designed for quality evaluation of instance Text spotting models.

PyTorch

For example, to evaluate text-spotting-0001 model on ICDAR2015 test dataset using PyTorch backend run:

python tools/test.py \
    --prob_threshold 0.8 \
    --dataset IC15TEST.json \
    --mean_pixel 123.675 116.28 103.53 \
    --std_pixel 58.395 57.12 57.375 \
    --rgb \
    --size 768 1280 \
  pytorch \
    --model configs/text-spotting-0001.json \
    --weights <path_to_checkpoint>.pth

Note: Use CUDA_VISIBLE_DEVICES environment variable to configure which (if any) GPUs to use during evaluation. If empty value is assigned, PyTorch uses CPU backend.

Demo

In order to see how trained model works using OpenVINO™ please refer to Text Spotting Python* Demo. Before running the demo you have to export trained model to IR. Please see below how to do that.

Export PyTorch Models to OpenVINO™

To run the model via OpenVINO™ one has to export PyTorch model to ONNX first and then convert to OpenVINO™ Internal Representation (IR) using Model Optimizer.

Model will be split into three parts:

• Text detector (Mask-RCNN like)
• Additional text recognition head
  • Text recognition encoder
  • Text recognition decoder

Export to ONNX*

The tools/convert_to_onnx.py script exports a given model to ONNX representation.

python tools/convert_to_onnx.py \
    --model configs/text-spotting-0001.json \
    --ckpt <path_to_checkpoint>.pth \
    --input_size 768 1280 \
    --show_flops \
    --output_folder export

Convert to IR

Conversion from ONNX model representation to OpenVINO™ IR is straightforward and handled by OpenVINO™ Model Optimizer. Please refer to Model Optimizer documentation for details on how it works.

1. text-spotting-0001-detector:

   mo.py \
       --model_name text-spotting-0001-detector \
       --input_shape="[1,3,768,1280],[1,3]" \
       --input=im_data,im_info \
       --mean_values="im_data[123.675,116.28,103.53]" \
       --scale_values="im_data[58.395000005673076,57.120000003655676,57.37500003220172],im_info[1]" \
       --output=boxes,scores,classes,raw_masks,text_features \
       --reverse_input_channels \
       --input_model export/detector.onnx

2. text-spotting-0001-encoder:

   mo.py \
       --model_name text-spotting-0001-encoder \
       --input_model export/encoder.onnx

3. text-spotting-0001-decoder:

   mo.py \
       --model_name text-spotting-0001-decoder \
       --input_model export/decoder.onnx