Missing `object_anchor` metadata for MBench-A object subset

[forLG]

Hi, thanks for releasing MBench and the evaluation code.

I encountered an issue when evaluating the MBench-A object consistency metrics:

• mbencha.entity.object_texture_consistency
• mbencha.entity.object_geometry_consistency

Both metrics require metadata.object_anchor, but the sample.json files in the currently released MBench-A object subset do not seem to contain this field.

Using the latest HuggingFace dataset snapshot, the downloaded sample.json contains:

{ 
  "sample_id": "sample_115_b7816bd8",
  "subset": "object",
  "caption": "...",
  "ground_truth_media": {
    "source_video": "source_video.mp4",
    "first_frame": "first_frame.png",
    "reference": "reference.png"
  }
}

but missing object_anchor. I checked all 100 samples under MBench-A/samples/object/*/sample.json, and found that all 100 are missing it.

However, the current evaluator requires this field. The object metrics specify contracts equivalent to:

object_texture_consistency:
  media.videos
  metadata.object_anchor

object_geometry_consistency:
  media.videos
  artifacts.da3
  metadata.object_anchor

Interestingly, the GitHub demo sample(sample_115_b7816bd8) includes an object_anchor, e.g.:

"object_anchor": {
  "center_x": 748.0,
  "center_y": 371.0,
  "width": 1280,
  "height": 720,
  "source": "spatial_subset_ranking"
}

Could you please clarify:

1. Are the object_anchor annotations for the full MBench-A object subset available somewhere?
2. Are these annotations supposed to be included in samples/object/<sample_id>/sample.json?
3. If they need to be generated separately, could you provide the corresponding annotation file or preprocessing script?
4. Which version of the object-anchor metadata was used for the official MBench-A leaderboard results?

Without these annotations, it seems difficult to reproduce the official Object Texture and Object Geometry evaluation faithfully, since generating anchors independently may result in different target-object selections and therefore non-comparable scores.

Thanks!