Frame Rate Hypothesis
We tested whether increasing the camera sampling rate from 6.7fps to 10fps improves cross-modal verification. It made things worse. Here's why.
The Hypothesis
Higher sampling rate → more camera frames → finer-grained direction data → improved direction agreement.
The moving-blob tracker samples at 150ms intervals (6.7fps), producing ~49 camera frames per 8-second run. Reducing the interval to 100ms (10fps) should yield ~65+ frames — a 33% increase in data density. Intuitively, more data should improve the direction agreement metric, which was ~65% at 6.7fps.
Method
| Parameter | Control (6.7fps) | Test (10fps) |
|---|---|---|
| Sampling interval | 150ms | 100ms |
| Tracker | moving-blob | moving-blob |
| Device | iPhone + desktop camera | Same |
| Runs | 60 (daylight) | 10 (daylight) |
Results
| Metric | 6.7fps | 10fps | Δ |
|---|---|---|---|
| Camera frames (avg) | 49 | 65 | +33% |
| Pass rate | 87% | 70% | -17% |
| Direction agreement | ~65% | 60% | -5% |
| Temporal alignment | 100% | 100% | — |
Higher frame rate decreased both pass rate and direction agreement. Temporal alignment remained at 100% — the coupling signal survived — but the direction signal degraded with faster sampling.
Why It Failed
At 150ms intervals, the centroid of moving blobs shifts by an average of 3–5 pixels between frames — enough displacement for a clean direction vector. At 100ms intervals, the displacement shrinks to 1.5–3 pixels.
The 1.5 pixel movement threshold acts as a low-pass filter: at 10fps, a larger fraction of true biological motion falls below the noise floor. The same motion that produced a clear signal at 150ms produces sub-threshold noise at 100ms.
Lesson: More data ≠ better data. The optimal sampling rate is determined by the signal amplitude, not the sensor capability. For a pixel-blob tracker with a 1.5px threshold, 150ms (~6.7fps) is near the optimal balance between temporal resolution and signal-to-noise ratio.
What We'll Do Differently
- Test before committing. A 10-run pre-test would have caught this without wasting a full batch.
- Adaptive threshold. The movement threshold (1.5px) should scale with the sampling interval: tighter intervals need proportionally relaxed thresholds.
- Report failures promptly. This report was written the day after the experiment. Future failures will be published within the same session.