How can I obtain the detailed semantic orientation annotations?

[ZhexiLuo]

Hello, thanks for the great work on SoFar!

I’m trying to use the semantic-orientation annotations in the dataset, but I’m confused about how to access the precise semantic orientation labels.

I downloaded OrienText300K.json, but the entries seem to contain only simplified orientation information, e.g.:

{
    "id": "6b4a9500cf154f06a50c96ae4fd500b6",
    "folder": "000-069",
    "point": "6b4a9500cf154f06a50c96ae4fd500b6.pt",
    "conversations": [
        {
            "from": "human",
            "value": "<point>\nfront"
        },
        {
            "from": "gpt",
            "value": ""
        }
    ],
    "rotation": [0, 0, -90],
    "direction": [-1, 0, 0]
}

In another issue, the authors mentioned that the provided rotation and direction values are mainly for visualization and not the actual fine-grained semantic orientation annotations.

My question:

How can I obtain the accurate semantic orientation labels used in the paper (e.g., the precise annotate directions)
Is there a separate file, script, or annotation process that provides the full semantic-orientation ground-truth?

Thanks in advance for your help!

[ZhexiLuo]

by the way, how to obtain the high-quality 15k subset of objaverse-LVIS according to paper?

We obtain the subset for OrienText300K from Objaverse-LVIS, which consists of approximately 46,000 3D objects with category annotations. The selection was based on the seven criteria mentioned in the main text. **Objects meeting all seven criteria formed the strict subset, comprising around 15k objects.** When including objects without textures and those of lower quality, the total increases to approximately 26k objects.