DATASET.md

Dataset preparation

Create a your-folder/datasets directory.

CUB

1. Download CUB dataset and unzip it in your-folder/datasets.
2. Download annotations from U-CMR using this link and unzip them in your-folder/datasets.

At the end you should obtain a your-folder/datasets directory organized as follows:

   .
   ├── CUB_200_2011/             # CUB original dataset
   └── UCMR_CUB_data/            # U-CMR annotations

Pascal3D+

1. Follow option A) or B) to have preprocessed data of Pascal3D+.
2. Download annotations of car class from U-CMR using this link and unzip them in your-folder/datasets. These data are used when testing on the car class to reproduce results of Table 2 in the paper.

A) Download preprocessed data

To obtain the already preprocessed data, please send an email to Alessandro Simoni stating:

1. Your name, title and affiliation
2. Your intended use of the data

B) Preprocess data yourself

Install Detectron2 (v0.3) in your environment running this line:

python -m pip install detectron2==0.3 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu102/torch1.5/index.html

1. Create PASCAL3D+ directory in your-folder/datasets.
2. Create PASCAL3D+/images directory in your-folder/datasets.
3. Download and unzip the Pascal3D+ dataset.
4. Copy and paste the content of Images folder of the unzipped dataset in your-folder/datasets/PASCAL3D+/images.
5. Copy and paste PASCAL/VOCdevkit/VOC2012/SegmentationClass folder of the unzipped dataset in your-folder/datasets/PASCAL3D+.
6. Copy and paste PASCAL/VOCdevkit/VOC2012/SegmentationObject folder of the unzipped dataset in your-folder/datasets/PASCAL3D+.
7. Download the annotations/masks folder and unzip it into your-folder/datasets/PASCAL3D+.

At the end you should obtain a your-folder/datasets/PASCAL3D+ directory organized as follows:

   .
   ├── annotations/             # folder with txt annotation files
   ├── images/                  # folder with Imagenet/PascalVOC image folders
   │   ├── aeroplane_imagenet/    
   │   │                  ├── first_image.jpg
   │   │                  ├── ...
   │   │                  └── last_image.jpg
   │   ├── aeroplane_pascal/             
   │   ├── ...             
   │   ├── tvmonitor_imagenet/            
   │   └── tvmonitor_pascal/                
   ├── masks_cmr/               # folder with CMR masks
   │          ├── aeroplane.mat
   │          ├── ...
   │          └── tvmonitor.mat
   ├── masks_hariharan/         # folder with Hariharan masks
   │          ├── 1
   │          │   ├── first_mask.png
   │          │   ├── ...
   │          │   └── last_mask.png
   │          ├── ...
   │          └── 20
   ├── SegmetationClass/        # folder with segmentation class masks for PascalVOC
   │   │             ├── first_image.jpg
   │   │             ├── ...
   │   │             └── last_image.jpg
   ├── SegmentationObject/      # folder with segmentation object masks for PascalVOC
   │   │             ├── first_image.jpg
   │   │             ├── ...
   │   │             └── last_image.jpg
   ├── aeroplane_val.txt
   ├── bicycle_val.txt
   ├── boat_val.txt
   ├── bottle_val.txt
   ├── bus_val.txt
   ├── car_val.txt
   ├── chair_val.txt
   ├── diningtable_val.txt
   ├── motorbike_val.txt
   ├── sofa_val.txt
   ├── train_val.txt
   ├── tvmonitor_val.txt
   ├── VOC_classes.txt
   └── VOC_mask_colors_RGB.txt

Run the preprocessing scripts as follows:

   python datasets/pascal3d/preprocess.py --dataset_dir your-folder/datasets/PASCAL3D+ --results_dir your-folder/datasets/PASCAL_final
   python datasets/pascal3d/split_train_val_test_VOC.py --orig_dataset_dir your-folder/datasets/PASCAL3D+ --new_dataset_dir your-folder/datasets/PASCAL_final

At the end you should obtain a your-folder/datasets/PASCAL_final directory organized as follows:

   .
   ├── aeroplane/                       # class folder
   │   ├── annotations/                 # class annotations folder
   │   │             ├── first_annot.yaml
   │   │             ├── ...
   │   │             └── last_annot.yaml
   │   ├── images/   
   │   │        ├── first_image.jpg
   │   │        ├── ...
   │   │        └── last_image.jpg       
   │   ├── masks/   
   │   │        ├── first_mask.png
   │   │        ├── ...
   │   │        └── last_mask.png         
   │   ├── masks_cmr/   
   │   │        ├── first_mask.png
   │   │        ├── ...
   │   │        └── last_mask.png          
   │   ├── masks_hariharan/  
   │   │        ├── first_mask.png
   │   │        ├── ...
   │   │        └── last_mask.png           
   │   ├── masks_pointrend/ 
   │   │        ├── first_mask.png
   │   │        ├── ...
   │   │        └── last_mask.png            
   │   ├── masks_VOC/       
   │   │        ├── first_mask.png
   │   │        ├── ...
   │   │        └── last_mask.png      
   │   ├── aeroplane_eval.txt            
   │   ├── aeroplane_test.txt       
   │   └── aeroplane_train.txt                    
   ├── bicycle/
   ├── boat/
   ├── bottle/
   ├── bus/
   ├── car/
   ├── chair/
   ├── diningtable/
   ├── motorbike/
   ├── sofa/
   ├── train/
   └── tvmonitor/