Datasets:

---

Kelly0510

/

FitAQA

Like 3

Hi, thank you for releasing FitAQA!

We would like to use the QEVD-FIT-COACH-Competition subset of temporal_grounding. Our goal is to run pose estimation and align per-frame keypoints with the annotated form-error intervals.

We have the QEVD-FIT-COACH-Competition-CVPR2025 release. Its feedbacks_long_range.json contains:

• long_range_video_file
• video_timestamps
• exercises
• exercise_start_timestamps
• exercise_end_timestamps

However, it does not contain feedbacks, feedback_timestamps, or is_transition, which the released preprocess.py requires. We can reconstruct the COACH and Benchmark clip boundaries using those fields, but cannot reproduce the Competition clips from our current release.

Could you please clarify:

1. Which exact Competition release did you use? Is it different from the CVPR2025 release? Where can we obtain the annotations required by preprocess.py?

2. Did Competition use the same preprocessing procedure as COACH/Benchmark, or a different method based on exercise start/end timestamps? If different, could you share the relevant code and clip naming/indexing rules?

3. Could you provide a clip-to-source mapping manifest for the Competition clips referenced in FitAQA? For example:
   dataset, video_name, source_video, start_frame, end_frame, output_fps
   Please also specify whether frame indices are zero-based and whether end_frame is inclusive. Source-video checksums would help verify the correct QEVD release.

4. Are the gold start/end times measured on the re-encoded clip's playback timeline? Was there any frame dropping, duplication, or resampling? How should these intervals be converted back to original-video frame indices?

A mapping manifest would allow us to align the annotations accurately even if the processed videos cannot be redistributed.

Thank you for your help!