Potential action_log annotation/timestamp inconsistencies in VideoCUA: Joplin/51533 and Brackets/91222

#4
by Gradygu3u - opened

Hi VideoCUA/GroundCUA team,

Thank you for releasing this dataset. While evaluating a video-only GUI trajectory extraction pipeline against VideoCUA, we manually inspected several disagreement cases and found two examples that may need annotation/timestamp verification in the original action_log.json.

I am reporting them here in case they indicate timestamp drift, missing intermediate actions, or action-log conversion ambiguity.

Case 1: Joplin / 51533

  • Dataset path: data/Joplin/51533/
  • Video: data/Joplin/51533/video/video.mp4
  • Action log: data/Joplin/51533/action_log.json
  • Task instruction: Organize research topics in numbered bullet points in Joplin
  • Raw action index: 1
  • Raw action type: CLICK
  • Timestamp: 5.000s
  • Params: {x: 515, y: 194, text: "Left", numClicks: 1}
  • Converted action: left_click @ 5.000s

During manual visual inspection, this logged click appears inconsistent with the visible operation around that timestamp. A nearby detected click around 2.000s also appears to be a different operation rather than the same action with a shifted timestamp.

Could you please verify whether the CLICK @ 5.000s annotation is correct?

Case 2: Brackets / 91222

  • Dataset path: data/Brackets/91222/
  • Video: data/Brackets/91222/video/video.mp4
  • Action log: data/Brackets/91222/action_log.json
  • Task instruction: Display the list of all available keyboard shortcuts from the help section.
  • Raw action index: 4
  • Raw action type: SCROLL
  • Timestamp: 11.928s
  • Params: {scrollX: 1, scrollY: 10}
  • Converted action: scroll down, amount=10 @ 11.928s
  • Cursor coordinate inherited from previous action in our conversion: [383, 108]

During manual visual inspection, this logged scroll also appears inconsistent with the visible operation around that timestamp. The closest detected action around 9.000s is a left_click, which seems to be a different neighboring operation rather than the same action.

Could you please verify whether the SCROLL @ 11.928s annotation is correct?

Attached evidence

I attached marked contact sheets for both cases:

  • Red marker: the original VideoCUA gold action timestamp/action.
  • Blue marker: the nearest detected candidate from our video-only pipeline.

These examples are present in the final converted trajectory we used, but the source of the gold action is the original VideoCUA action_log.json.

Thanks again for the dataset. We are trying to use VideoCUA as a reference benchmark for video-to-GUI-trajectory extraction, so confirming these cases would help us understand whether they are annotation issues, timestamp alignment issues, or expected behavior.

Sign up or log in to comment