Task

Various objects are one-by-one placed into the standard setup at different positions and orientations. For each of these frames, the object in the scene must be identified.

Compared to the other challenges where complex object properties such as its shape, texture or position are to be reconstructed the answer in each frame here is just one out of eleven object identifiers. However, due to the varying placements in the scene, this challenge is still far from being easy.

The objects were scaled to equal surface area to prevent classifying by the size of the object, i.e. amount of reflected light. This approach is not necessarily perfect, as very compact or concave objects appear smaller when scaled by their surface area, but it is a decent compromise.

Dataset

The dataset contains the eleven following objects:

Each object is placed at eight different positions in the scene, with a varying orientation. So for a total of 88 frames the correct object must be identified.

Ground truth geometric information is given for all objects. It is therefore a classification of known objects.

The spoon is contained twice in the dataset, where the only difference is the material. This tests object classification purely by material and not shape.

Database

Models

Submission

The submission should consist of a single json file, that stores the classification information about each frame of the dataset. Fuzzy classification is supported; just give multiple weights as follows:

[
    {
        "frame": 0,
        "class": {
            "Whale": 0.9147317973627079,
            "Cat": 0.0729247274133686,
            "SpoonDiffuse": 0.012343475223923438
        }
    },
    {
        "frame": 1,
        "class": "Pan"
    },
    {
        "frame": 2,
        "class": {
            "Plant": 0.8174477797852516,
            "SpoonSpecular": 0.18255222021474835
        }
    },
    {
        "frame": 3,
        "class": "Parallelepiped"
    },
    ... etc
]

The weights for one frame must always add up to 1.