Material Classification

Authors

1. Truong Phuc Anh - 14520040@gm.uit.edu.vn

Requirements

1. Python 3.6

2. Matlab

Setup

1. Run ./setup/install_python.cmd to install python 3.

2. Run command python in command line to check whether install sucess or not.

3. Run ./setup/install_packages.cmd to install python's packages.

4. Run python ./setup/test/tensorflow_test.py to validate your tensorflow installation. If "Hello, TensorFlow!" is printed, your installation is success.

5. Run python ./setup/test/keras_test.py to validate your keras installation. If "Hello Keras!" is printed, your installation is success.

Dataset

Download 2 datasets using link below, extra it in ./src/DS_/original.

Re-format dataset into ./src/DS_/original/

1. GTOS dataset: http://eceweb1.rutgers.edu/vision/gts/download.html

2. FMD dataset: https://people.csail.mit.edu/celiu/CVPR2010/FMD/

Get texture images for a dataset

1. Make sure you have original dataset in folder ./DS_<dataset_name>/data/original.

2. Run the following command.

python ./src/tools/create_data_folders.py <dataset_name> texture
python ./src/get_data_filenames.py <dataset_name> original

1. Open Matlab.exe, change working folder to ./src.

2. In get_texture.m edit txt file name into <dataset_name>_original.txt.

3. Run get_texture.m.

4. Open folder ./DS_<dataset_name>/data/texture to check

Get edges images

1. Make sure you have original dataset in folder ./DS_<dataset_name>/data/original.

2. Run the following command.

python ./src/local-feature-extractor/extract_candy_edge.py <dataset_name> 

Train and test with SVM

1. Make sure you have original, edges and texture dataset in folder ./DS_<dataset_name>/data.

2. Configs your own SVM kernel parameter in file ./config/models_config.csv

3. Run the following command.

./src/run-FMD.cmd
./src/run-GTOS.cmd

Fine-tuning with VGG16

1. Only re-train the last layer. Run the following commands:

python .\src\vgg16_experiment\finetune.py GTOS 1 39
python .\src\vgg16_experiment\finetune.py GTOS 2 39
python .\src\vgg16_experiment\finetune.py GTOS 3 39
python .\src\vgg16_experiment\finetune.py GTOS 4 39
python .\src\vgg16_experiment\finetune.py GTOS 5 39

python .\src\vgg16_experiment\finetune.py FMD 1 10
python .\src\vgg16_experiment\finetune.py FMD 2 10
python .\src\vgg16_experiment\finetune.py FMD 3 10
python .\src\vgg16_experiment\finetune.py FMD 4 10
python .\src\vgg16_experiment\finetune.py FMD 5 10

2. Re-train all layers. Run the following commands

python .\src\vgg16_experiment\finetune_full.py GTOS 1 39
python .\src\vgg16_experiment\finetune_full.py GTOS 2 39
python .\src\vgg16_experiment\finetune_full.py GTOS 3 39
python .\src\vgg16_experiment\finetune_full.py GTOS 4 39
python .\src\vgg16_experiment\finetune_full.py GTOS 5 39

python .\src\vgg16_experiment\finetune_full.py FMD 1 10
python .\src\vgg16_experiment\finetune_full.py FMD 2 10
python .\src\vgg16_experiment\finetune_full.py FMD 3 10
python .\src\vgg16_experiment\finetune_full.py FMD 4 10
python .\src\vgg16_experiment\finetune_full.py FMD 5 10

How to view result?

Results are stored in ./<dataset_name>/result/test

Including: accuracy, miss sample, classifiers, confusion matrix, etc.