Code repository for THOR presented in Persistent Homology Meets Object Unity: Object Recognition in Clutter.

The UW Indoor Scenes (UW-IS) Occluded Dataset proposed in the above paper for systematically evaluating object recognition methods under varying environmental conditions can be found here. This dataset, which is recorded using commodity hardware, consists of two different indoor environments, multiple lighting conditions, and multiple degrees of clutter.

• Panda3D
• Open3D
• persim 0.3.1
• scikit-learn
• Keras
• Platform: This code has been tested on Ubuntu 18.04 (except synthetic data generation using Panda3D, which is done on a computer running Windows 10).

• Training:

  1. Follow instructions from here to install Panda3D for synthetic data generation. Create a new folder syndata in the directory where Panda3D is installed and place generateSyntheticData.py from the training folder in this repository into the syndata folder. Create subfolders models and data inside the syndata folder. Within models, create subfolders for all objects and place respective object meshes and texture maps inside them. Obtain synthetic depth images from the object meshes using Panda3D using the following command.

     python generateSyntheticData.py --obj_name <obj_name> --h <h> --p <p> --r <r>

     <obj_name> is the name of object for which data is to be generated, and the parameters h,p, and r are set to reorient the object mesh as required (details in the paper) before rendering. This command will create synthetic depth (and RGB) images for the object under a subfolder <obj_name> inside the data folder.

  2. From within the THOR directory run the following to generate point clouds corresponding to all the generated depth images.

     python3 training/getPCDsFromSyntheticData.py --data_path <path_to_data_folder_from_step_i>

  3. From within the THOR directory run the following to perform view normalization on the generated point clouds.

     python3 training/saveAllViewNormalizedPCDs.py --data_path <path_to_data_folder_from_step_i>

  4. From within the THOR directory run the following to generate Persistence Images (PIs) for the TOPS descriptor of all the point clouds.

     python3 training/computePIsFromViewNormalizedPCDs.py --data_path <path_to_data_folder_from_step_i>

     A subfolder named libpis containing all the PIs will be generated inside the training folder .

  5. Run the following to train an SVM library using the TOPS descriptors obtained from the computed PIs. (Add the path to the data folder from step i in trainSVMLibrary.sh as indicated).

     cd training
     sh trainSVMLibrary.sh

     Alternatively, to train an MLP library run the following. (Add the path to the data folder from step i in trainMLPLibrary.sh as indicated).

     cd training
     sh trainMLPLibrary.sh

     A folder librarymodels will be created inside the training directory and trained models will be stored in it.

• Testing on the UW-IS Occluded Dataset:

  1. Download the UWISOccludedDataset.zip from here and unzip it. Place reogranizeUWISOccluded.sh inside the UWISOccludedDataset folder and run the following from within that folder.

     sh reorganizeUWISOccluded.sh

  2. Run the following to test THOR

     • Using an SVM library:

        cd testing
        sh testUWISOccludedSVMLibrary.sh

       Note that in the testUWISOccludedSVMLibrary.sh script <environment_name> must be replaced with one of warehouse,lounge, or both as desired. Similarly <category_name> can be kitchen, food, tools or all; <separation> can be level1, level2, level3 or alllevels; <light> can be 1,2, or both. Also provide the path to the UW-IS Occluded dataset folder from the previous step, and the path to the folder containing saved SVM models.

       A subfolder named predictions is created in the testing folder and predictions for every video will be saved as a .txt file. Corresponding ground truth will be saved in a newly created groundtruth subfolder.

     • Using an MLP library:

        cd testing
        sh testUWISOccludedMLPLibrary.sh

       Note that <environment_name>, <category_name>, <separation>, and <light> are to be replaced as described above. The path to the dataset folder and the folder containing trained MLP models must also be provided. As in the SVM case, a subfolder named predictions is created in the testing folder and predictions for every video will be saved as a .txt file. Corresponding ground truth will be saved in a newly created groundtruth subfolder.