Notice that we use pyrender with OSMesa support, you may need to install OSMesa after running pip install pyrender, more details can be found here.

MinkowskiEngine append its package path into sys.path (a.k.a., PYTHONPATH), which includes a module named utils. In order not to get messed with our own utils package, you should import MinkowskiEngine after importing utils.

Download ShapeNet v2 dataset and modify the shapenet_root key in config/config.yaml to point to the location of the dataset.

To train on synthetic ShapeNet objects that appear in NOCS REAL275, run:

python train.py category=bottle,bowl,camera,can,laptop,mug -m

For laptops, an auxiliary segmentation is needed to ensure a unique pose. Please refer to Auxiliary Segmentation for Laptops/

To train on synthetic ShapeNet objects that appear in SUN RGB-D, run:

python train.py category=bathtub,bed,bookshelf,chair,sofa,table -m

For Laptops, geometry alone cannot determine the pose unambiguously, we rely on an auxiliary segmentation network that segments out the lid and the keyboard base.

To train the segmenter network, first download our Blender physically rendered laptop images from Google Drive and place it under data/laptop. Then run the following command:

python train_laptop_aux.py

First download the detection priors from Google Drive, which is used for evaluation with instance segmentation or bounding box masks. Put the directory under data/nocs_seg.

Then download RGB-D images from NOCS REAL275 dataset and put it under data/nocs.

Place (pre-)trained models under checkpoints.

First save inference outputs:

python nocs/inference.py --adaptive_voting

Then evaluate mAP:

python nocs/eval.py | tee nocs/map.txt

First save inference outputs with bounding box mask enabled:

python nocs/inference.py --bbox_mask --adaptive_voting

Then evaluate mAP:

python nocs/eval.py | tee nocs/map_bbox.txt

For this task, due to the memory limitation, we use the regression-based network. You can go through the process by running the jupyter notebook nocs/zero_shot.ipynb.

We follow the same data preparation process as in VoteNet. You need to first download SUNRGBD v2 data (SUNRGBD.zip, SUNRGBDMeta2DBB_v2.mat, SUNRGBDMeta3DBB_v2.mat) and the toolkits (SUNRGBDtoolbox.zip). Move all the downloaded files under data/OFFICIAL_SUNRGBD. Unzip the zip files.

Download the prepared extra data for SUN RGB-D from Google Drive, and move it under data/sunrgbd_extra. Unzip the zip files.

First save inference outputs:

python sunrgbd/inference.py

Then evaluate mAP:

python sunrgbd/eval.py | tee sunrgbd/map.txt

To train on custom objects, it is necessary to understand some parameters in configuration files.

• up_sym: Whether the objects look like a cylinder from up to bottom (e.g., bottles). This is to ensure the voting target is unambiguous.
• right_sym: Whether the objects look like a cylinder from left to right (e.g., rolls). This is to ensure the voting target is unambiguous.
• regress_right: Whether to predict the right axis. Some symmetric objects only have a up axis well defined (e.g., bowls, bottles), while some do not (e.g., laptops, mugs).
• z_right: Whether the objects are placed such that the right axis is [0, 0, 1] (default: [1, 0, 0]).

Next, we need to know the scale_range (used for data augmentation, control possible object scales along the diagonal), vote_range (the range for center voting targets $\mu,\nu$), and scale_mean (the average 3D scale, used for scale voting). To generate them, you may refer to gen_stats.py.

After you prepare the necessary configurations and voting statistics, you can write your own configuration file similar to that in config/category, and then run train.py.