Does the mask_id in VPData mean anything?

[billcao2000]

Thank you for open-sourcing this work and the dataset! I noticed that the mask annotation files in VPData contain segmentation results for multiple masks, distinguished by the mask_id field. How was this field generated? Does each numerical value correspond to a real-world category, similar to how detection models assign category IDs?

[yxbian23]

Thank you for open-sourcing this work and the dataset! I noticed that the mask annotation files in VPData contain segmentation results for multiple masks, distinguished by the mask_id field. How was this field generated? Does each numerical value correspond to a real-world category, similar to how detection models assign category IDs?

Thanks for the question!

The mask_id values are simply intermediate identifiers produced by the Recognize-Anything model during data construction.
Each video clip may contain multiple separate mask sequences, and each sequence is assigned a unique mask_id within that clip.

However, these IDs do not correspond to semantic categories and are not consistent across clips — i.e., mask_id = 1 in one video does not relate to mask_id = 1 in another video.

So in short:

• Each mask_id just indexes one mask track within a single clip
• No cross-clip meaning or category mapping is implied