SyntheT-Less

Official repository for the SyntheT-Less dataset

used in On Object Symmetries and 6D Pose Estimation from Images, G. Pitteri*, M. Ramamonjisoa*, S. Ilic and V. Lepetit

Download our data

The data is available here (9.6GB zip file).

Dataset structure

Here is how the dataset is organised:

multi_objects/
|__images
    |__ img_xxxxx.png
|__contours/
    |__ xxxxx.png
|__depth/
    |__ xxxxx.png                         
|__normals
    |__ xxxxx.png
|__mask
    |__ xxxxx.png
|__instances
    |__ xxxxx.png --> uint8 image containing object labels from 1 to 9 (background pixels are annotated with 0) 
|__gt_poses/
    |__ tmp_xxxxx.json --> contains a dictionnary of object poses with keys annotated from 0 to 8 
                           (There is a 1-shift from object instances indices, sorry about that!)
        {
         "0": {"Euler": [theta_x, theta_y, theta_z],
               "T": [Tx, Ty, Tz],
               "type": object type from T-LESS (from 1 to 30)},              
         ...
                  
         "8": {"Euler": [theta_x, theta_y, theta_z],
               "T": [Tx, Ty, Tz],
               "type": object type from T-LESS (from 1 to 30)},
         "Lamp": {"phi": lamp_phi, 
                  "theta": lamp_theta,
                  "strength": lamp_strength}
	  "Ambient Light": ambient occlusion level (0 to 1)
	  "Table": {"distance": distance in meters,
		    "Euler": [theta_table_x, theta_table_y, theta_table_z]}
         }   

Data Generation

Requirements

1. Download the DTD dataset

Download the Describable Textures Dataset (DTD) and set its folder using:

cd DTD && \
wget https://www.robots.ox.ac.uk/~vgg/data/dtd/download/dtd-r1.0.1.tar.gz && \
tar -xf dtd-r1.0.1.tar.gz --strip-components=1 && cd .. &&\ 
rm -rf DTD/dtd-r1.0.1.tar.gz

2. Download the T-LESS CAD models

Download the T-LESS CAD models:

wget http://ptak.felk.cvut.cz/darwin/t-less/v2/t-less_v2_models_cad.zip && \
unzip t-less_v2_models_cad.zip && \  
rm -rf t-less_v2_models_cad.zip

3. Install Blender

Download Blender and save it to the Blender/ path. using the following commands (skip this step if you have previously installed Blender).

mv ABSOLUTE_PATH_WHERE_FILE_WAS_DOWNLOADED.tar.bz2 ROOT_OF_THIS_DIRECTORY 
bzip2 -d blender-2.80-linux-glibc217-x86_64.tar.bz2
mkdir blender && tar -xf blender-2.80-linux-glibc217-x86_64.tar -C blender/ --strip-components=1
rm -rf blender-2.80-linux-glibc217-x86_64.tar

Run the generation code

Make sure you have installed the following requirements:

• Blender (in previous step)
• OpenCV
• OpenEXR
• imageio
• skimage
• multiprocessing
• imath

Most of them can be installed using pip install "package_name"

Then run:

python3 call_blender_multi.py --blender_path $BLENDER_PATH \
			      --dtd-rootdir DTD \
			      --models_path models_cad\
			      --cpus $NUM_CPUS --size $DATASET_SIZE --cuda $CUDA_DEVICE

Variables:

• $BLENDER_PATH: Path to the directory containing your Blender executable (use blender/ if installed with step 3)
• $NUM_CPUS: Number of processes to run data generation in parallel. We recommand setting it to your machine CPUs number.
• $DATASET_SIZE: Aimed dataset size
• $CUDA_DEVICE: GPU ID for data generation

Call python3 call_blender_multi.py --help for help on other parameters.

Citation

If you use our data generation code or already generated data, please cite our paper:

@article{pitteri2019threedv, 
 Title = {On Object Symmetries and 6D Pose Estimation from Images}, 
 Author = {G. Pitteri and M. Ramamonjisoa and S. Ilic and V. Lepetit}, 
 Journal = {International Conference on 3D Vision}, 
 Year = {2019}
 }