Temporal Pooling Dashboard

Exploring the Temporal Integration of Video Quality: How algorithmic metrics (CVVDP) align with human perception (Subjective JOD) across varying quality profiles.

1. Calibration (Human vs. Metric)

Subjective vs. CVVDP (Varying)

This plot shows subjective and cvvdp scores for every unique 3-chunk combination, e.g. 1080p30-720p60-1080p30. There are 150 such combinations.

Subjective vs. CVVDP (Fixed)

The baseline comparison using single-quality videos (no fps and resolution changes).

2. Subjective Integration (Human Perception)

While the scaled JOD for Fixed videos is reasonable, the values for Varying videos are currently unstable for several scenes:

Bistro: -31 to -310
Gallery: -99 to -415
Sibenik: -229 to -1238

Only the "All" row, School, and Makeway show reasonable values (could still be negative). This instability occurs when the pw_scale_table has insufficient observer data or high observer disagreement for a specific scene. With 150 unique combinations (twice as many as previous experiments because we add reference videos), the current participant count is insufficient for robust scaling in these complex scenes. I'll recruit 10 more participants for the varying video experiment (estimated 1 hour per session) to resolve these scaling outliers.

Next Steps: We plan to recruit 10 more participants for the varying video experiment (estimated 1 hour per session) to resolve these scaling outliers.