pengjk0-dotcom / TennisProfile

# TennisProfile
**Reliability-Aware Groundstroke Profiling from Monocular 3D Human Recovery**
**Accepted at the MoCha workshop at ECCV 2026.** TennisProfile is a reproducible research project that transforms monocular 3D human-recovery outputs into *stroke-specific, expert-referenced* tennis motion profiles, with explicit attention to evidence reliability.
**Visual overview · Poster overview · Key findings · Reproduce · Latest release · Cite**
## Visual overview
TennisProfile connects reviewed stroke events, 3D body motion, and expert-reference comparisons. The project studies *what differs*, *where in the swing it differs*, and *how strongly the evidence supports the comparison*.
### Project at a glance
| Input and method | Research output |
| --- | --- |
| Monocular RGB groundstroke clips | Reviewed swing-start, ball-contact, and swing-end events |
| HMR2 human mesh recovery | Joint-motion representation and 11 scalar motion features |
| Five contact-anchored phases | Six-variable × five-phase deviation maps |
| Stroke-specific expert reference | Standardized feature deviations and reliability-aware interpretation |
## Evaluation cohort
**36 subjects · 72 reviewed strokes · 2 stroke types**
- 18 expert-reference subjects, 3 held-out experts, and 15 beginners.
- One flat forehand and one one-handed backhand per retained subject.
- The fixed selection was made independently of final feature values; see the cohort audit.
- The retained cohort is right-handed and limited in setting; it must not be presented as validation across every stroke, handedness, or in-the-wild camera view.
## Key findings
These are **study-specific descriptive results**, not product accuracy claims:
| Evaluation | Observed result | Evidence |
| --- | --- | --- |
| Mean absolute expert-reference deviation | **0.49 SD** held-out experts vs **0.76 SD** beginners | Robustness summary |
| Feature values outside the expert reference min–max | **0.0%** held-out experts vs **9.7%** beginners | Robustness summary |
| External cross-source comparison | Evaluated on **20 SportsPose tennis sequences from 4 subjects** | SportsPose validation |
The held-out expert group contains only three subjects; these comparisons should not be read as a generalizable classifier benchmark. The camera-ready robustness analysis also reports **no two-sided group-feature discoveries at BH q The GitHub companion is an accessible, reformatted version of the poster content; it is not an additional experiment or a replacement for the original high-resolution presentation file.
## Reproducibility
The repository makes available only material that can be redistributed: reviewed event tables, fixed cohort splits, derived motion features, group summaries, robustness results, figure assets, analysis scripts, and recorded environments.
**Verify the released tables** (Python standard library only):
 [code] 
**Reproduce analyses from included feature tables:**
 [code] 
Omit `--skip-sportspose` and set `SPORTSPOSE_ROOT