project

Project (2018-2019)

Installation

1. Python 3.6
2. TensorFlow (1.11)
3. Keras (2.2.4)
4. keras-vis (last version)
   • pip install git+https://github.com/raghakot/keras-vis.git

Datasets

CMU-Cornell iCoseg Dataset

1. Look at some groups of images: http://chenlab.ece.cornell.edu/projects/touch-coseg/iCoseg_dataset.pdf;
2. Download the dataset from: http://chenlab.ece.cornell.edu/downloads.html;
3. Move dataset folders inside the ./datasets/;
4. Copy 80 images and their ground truths.
   List of images to be copied: ./datasets/icoseg/subset_80/img_list.txt.
   ./datasets/icoseg/images/... -> ./datasets/icoseg/subset_80/images
./datasets/icoseg/ground_truth/... -> ./datasets/icoseg/subset_80/grund_truth

Folder structure:

.
├── ...
├── datasets
│   ├── icoseg
│   │   ├── ground_truth
│   │   ├── image_download
│   │   ├── images
│   │   ├── images_subset5_cvpr10
│   │   ├── subset_80
│   │   │   ├── ground_truth
│   │   │   ├── images
│   │   │   ├── img_list.txt
│   │   │   ├── train.txt
│   │   │   ├── val.txt
│   │   │   └── test.txt
│   │   ├── readme.txt
│   │   └── ...
│   └── ...
└── ...

Inria Aerial Image Labeling Dataset

1. Download the dataset from: https://project.inria.fr/aerialimagelabeling/download/;
2. Select a subset of images;
3. Use ./utils/tiff_to_png.ipynb to change img format.

The final structure of ./datasets:

.
├── datasets
│   ├── icoseg
│   │   ├── subset_80
│   │   └── ...
│   ├── inria_aerial
│   │   └── subset_chicago
│   │       ├── test
|   |       |   ├── test_img_from_patches           #imgs after merging 224x224 img patches(_h_w)    
|   |       |   ├── test_img_patches                #512x512 img patches(+img_h+img_w) for unet
|   |       |   ├── test_img_patches_for_sal        #224x224 img patches(+img_h+img_w), don't use
|   |       |   ├── test_mask_from_patches          #masks after merging 224x224 mask patches(_h_w)
|   |       |   ├── test_mask_patches               #512x512 mask patches(+img_h+img_w) for unet
|   |       |   ├── test_mask_patches_for_sal       #224x224 mask patches(+img_h+img_w), don't use
|   |       |   ├── test_None_sal_from_patches      #images after merging 512x512 patches (None)
|   |       |   ├── test_None_sal_patches           #512x512 vgg16_sal_map_None patches for unet
|   |       |   ├── test_None_sal_patches_224_224   #224x224 result patches from vgg16_sal_map_None
|   |       |   ├── test_guided_sal_from_patches    #images after merging 512x512 patches (guided)
|   |       |   ├── test_guided_sal_patches         #512x512 vgg16_sal_map_guided patches for unet
|   |       |   └── test_guided_sal_patches_224_224 #224x224 result patches from vgg16_sal_map_guided
│   │       ├── train
|   |       |   └── ...
│   │       └── val
|   |           └── ...
│   ├── playground
│   └── unseen_categories
│       └── test
└── ...

References

1. Visualizing and Understanding Convolutional Networks (Zeiler and Fergus, 2013)
2. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps (Simonyan et al., 2014)
3. Striving for Simplicity: The All Convolutional Net (Springenberg et al., 2015)
4. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization (Selvaraju et al., 2016)
5. iCoseg: Interactive Co-segmentation with Intelligent Scribble Guidance (Dhruv Batra et al, 2010)
6. U-Net: Convolutional Networks for Biomedical Image Segmentation (Ronneberger et al., 2015)
7. keras-vis for neural network visualization
8. https://machinelearningmastery.com/use-pre-trained-vgg-model-classify-objects-photographs/