HoHo 5k Subset

This is a living dataset. Today, we provide 4316 samples for training, and 175 for validation and hold back an additional 1072 for computing the private and public leaderboards. Additional, we intend to continue releasing training data throughout the challenge and beyond. The data take the following form:

Features({
        "order_id": Value(dtype="string"),
        # inputs 
        "K": Sequence(Array2D(dtype="float32", shape=(3, 3))),
        "R": Sequence(Array2D(dtype="float32", shape=(3, 3))),
        "t": Sequence(Sequence(Value(dtype="float32"), length=(3))), # in centimeters
        "gestalt": Sequence(Image()), 
        "ade20k": Sequence(Image()),  
        "depthcm": Sequence(Image()),  # in centimeters
        # result of Colmap reconstruction loaded in named tuples
        # More on format: https://github.com/colmap/colmap/blob/main/scripts/python/read_write_model.py#L47
        "images": Dict(namedtuple("Image", ["id", "qvec", "tvec", "camera_id", "name", "xys", "point3D_ids"])),
        "points3d": Dict(namedtuple( "Point3D", ["id", "xyz", "rgb", "error", "image_ids", "point2D_idxs"])),
        "cameras": Dict(namedtuple("Camera", ["id", "model", "width", "height", "params"])),
 
        # side info during training
        "mesh_vertices": Sequence(Sequence(Value(dtype="float32"), length=3)),
        "mesh_faces": Sequence(Sequence(Value(dtype="int64"))),
        "face_semantics": Sequence(Value(dtype='int64')),
        "edge_semantics": Sequence(Value(dtype='int64')),

        # targets
        "wf_vertices": Sequence(Sequence(Value(dtype="float32"), length=3)),  # in centimeters
        "wf_edges": Sequence(Sequence(Value(dtype="int64"), length=2)),
   })

These data were gathered over the course of several years throughout the United States from a variety of smart phone and camera platforms. Each training sample/scene consists of a set of posed image features (segmentation, depth, etc.) and a sparse point cloud as input, and a sparse wire frame (3D embedded graph) with semantically tagged edges as the target. Additionally a mesh with semantically tagged faces is provided for each scene durning training. In order to preserve privacy, original images are not provided.

Note: the test distribution is not guaranteed to match the training set.

Sample visualizations

Two visualizations of houses below are interactive! Grab one with your mouse and rotate!

The roofs below are interactive as well!

Additional notes on data

Depth

The depthcm is a result of running monocular depth model, and it is not ground truth by no means. If you need to have a GT depth, you can render the GT mesh in the training set using mesh_faces and mesh_vertices. The semi-sparse depth from the Colmap reconstructions with dense features, available in points3d is much more accurate, than depthcm.

At the inference time, mesh_faces is not available, so you can use only depthcm and colmap point cloud from points3d

Segmentation

You have two segmentations available. gestalt is domain specific model, which "sees-through-occlusions" and provides a detailed information about house parts. See the list of classes in "Dataset" section in the navigation bar.

ade20k is a standard ADE20K segmentation model (specifically, shi-labs/oneformer_ade20k_swin_large).

Organizers

Jack Langerman (Hover), Dmytro Mishkin (CTU in Prague / Hover), Ilke Demir (Intel), Hanzhi Chen (TUM), Daoyi Gao (TUM), Caner Korkmaz (ICL), Tolga Birdal (ICL)

Timeline

Competition Released: March 14, 2024
Entry Deadline: May 28, 2024
Team Merging: May 28, 2024
Final Solution Submission: June 4, 2024
Writeup Deadline: June 11, 2024

Prizes

1st Place: $10,000
2nd Place: $7,000
3rd Place: $5,000
Additional Prizes: $3,000

Please see the Competition Rules for additional information.

Cite

@misc{Langerman_Korkmaz_Chen_Gao_Mishkin_Birdal2024, 
        title={S23DR Competition at 1st Workshop on Urban Scene Modeling @ CVPR 2024}, 
        url={usm3d.github.io},
        howpublished = {\url{https://huggingface.co/usm3d}},
        year={2024},
        author={Langerman, Jack and Korkmaz, Caner and Chen, Hanzhi and Gao, Daoyi Mishkin, Dmytro and Birdal, Tolga}
    }

Datasets:

usm3d
/

hoho-train-set

Acknowledge license to accept the repository