docs/texas_ground_truth.md

OSU vs Texas Ground Truth - Play Detection Analysis

This document tracks observations, issues, and ground truth for the Texas video analysis.

Video Details

• Video: OSU vs Texas 01.10.25.mkv
• Note: This video has Variable Frame Rate (VFR) encoding - see texas_video_vfr_issue.md for details

---

Known Issues

1. Opening Kickoff False Positive ⚠️

Issue	Details
Real kickoff starts	~7:21 in video
Problem	Strange detection of play clock before 7:21 causing false positive reading
Status	Left unfixed - couldn't resolve without causing regression on other videos
Priority	Low - edge case, revisit later

2. Variable Frame Rate Timing Issues ⚠️

Issue	Details
Problem	VFR video caused timing/padding issues
Symptom	Padding 2 seconds before was sometimes actually 4 seconds or 6 seconds before
Impact	Play clipping ffmpeg logic not well set up for VFR videos
Note	4 seconds padding before may actually be preferable to 2 seconds
Status	Fixed: default padding for normal plays changed from 2s to 4s
Documentation	See texas_video_vfr_issue.md for technical details

3. FLAG Play Merge Duplication ✅

Issue	Details
Problem	Flag plays overlap with regular plays before them
Symptom	Video shows end of regular play, then flag for a few seconds, then skips BACK in time to where FLAG first appeared
Fix	Clip generation now merges overlapping plays into single segments (see _compute_mergeable_segments in ffmpeg_ops.py). When play A (x to x+10) and FLAG B (x+7 to x+15) overlap, we create one clip (x to x+15) instead of two—no duplicate footage, full context preserved
Status	Fixed

4. Special Play Max Duration Too Long ✅

Issue	Details
Previous value	10 seconds max for special plays (when clock reset happens)
Fix applied	Reduced to 7 seconds
Condition	Special plays end when whichever happens first: (1) scorebug/playclock disappears, or (2) max duration elapsed
Rationale	10 seconds was too long for most special plays (punts, FGs, XPs)

---

Positive Results ✅

• No significant missed plays detected - core play detection appears to be working well for this video

---

TODO Items

1. Investigate FLAG merge duplication issue (high priority)
2. Reduce special play max duration from 10s to 7s
3. Increase pre-play buffer from 2s to 4s for normal plays
4. Investigate VFR handling in ffmpeg clip extraction logic
5. Investigate opening kickoff false positive (low priority)

---

Comparison with Other Videos

Metric	Tennessee	Oregon	Texas
Opening kickoff	✅ Clean	✅ Clean	⚠️ False positive
FLAG merge	✅ Working	✅ Working	❌ Duplicating
VFR issues	N/A	N/A	⚠️ Present
Missed plays	~1 acceptable	3 missing	✅ None significant

---

Notes for Future Analysis

• This video is a good test case for VFR handling improvements
• FLAG merge logic needs to be more robust against timestamp inconsistencies
• The opening kickoff issue suggests the detection may be sensitive to pre-game content