bhargavaurala / accessmath-textboxes

A fork of the TextBoxes project for finetuning and deploying a neural network for detection of handwritten whiteboard content in lecture videos

accessmath-textboxes

This project is meant as a helper for our main work for detection of handwritten whiteboard content in lecture videos. It is a fork of TextBoxes.

Setting up Textboxes

1. Clone this repository. We will call the clone directory $CAFFE_ROOT

git clone https://github.com/bhargavaurala/accessmath-textboxes.git

cd accessmath-textboxes

2. Edit the makefile configuration file Makefile.config according to your system needs. Refer Caffe installation instructions for details about dependencies. Make sure that the python wrapper dependencies are installed since we need that for this project. This code has been tested on Ubuntu 14.04.

cp Makefile.config.example Makefile.config

mkdir build

3. Build caffe, caffe-python and test if build went correctly.

make -j8

make py

export PYTHONPATH=$PYTHONPATH:$CAFFE_ROOT/python

python -c "import caffe"

Download

1. Models trained on ICDAR 2013: Dropbox link BaiduYun link
2. Fully convolutional reduced (atrous) VGGNet: Dropbox link BaiduYun link
3. Compiled mex file for evaluation(for multi-scale test evaluation: evaluation_nms.m): Dropbox link BaiduYun link
4. Frame version of the AccessMath dataset from here. Download the 3-part zip archive and extract into a folder called AccessMathVOC and place in AccessMath-ICFHR18 project root.

export AM_DATA_DIR=/path/to/AccessMathVOC

Generate training and validation LMDBs.

1. cd $CAFFE_ROOT/data/AccessMath
2. ./create_data.sh
3. This will create train, validation and test LMDBs in $AM_DATA_DIR/AccessMath/lmdb

Train

1. In models/VGGNet/text/longer_conv_300x300/ Modify data_param in the first layer (data) in train.prototxt and test.prototxt as shown below

  data_param {
    source: "/path/to/AccessMathVOC/AccessMath/lmdb/AccessMath_train_lmdb"
    batch_size: 32
    backend: LMDB
  }

2. Use cd $CAFFE_ROOT/build/tools ./caffe train_net -iterations 10000 -solver models/VGGNet/text/longer_conv_300x300/solver.prototxt -weights /path/to/model_trained_on_icdar2013
3. You should see around 77.5% as the final validation performance.
4. Transfer the model to models/text_detection in the AccessMath root folder

accessmath-textboxes's Issues

fatal error: curand.h: No such file or directory

I meet a problem that "fatal error: curand.h: No such file or directory" when i install caffe and set use_cudnn =1.