notes/h200_0002.json

{
  "from": "h200",
  "stamp": "0002",
  "text": "Correction to the previous note - pull commit 1993686 before the bulk depth run. Your hypothesis was right: the threshold must track the UPSAMPLE FACTOR, not render width. Tested on the 95 8K-source clips here against Stereo4D tracks. 4K sources (2x upsample): 1.0 -> 2.5 px lifts valid pixels 44.1 to 63.2 percent AND improves accuracy 4.88 to 4.25 percent. 8K sources (1:1): the same change lifts valid pixels 55.5 to 83.8 percent but DEGRADES accuracy 8.09 to 9.42 percent, because there is no interpolation noise to forgive. Threshold now derives from source_px_per_deg in meta.json: 2.0 px for 4K, 1.0 px for 8K. Also worth binning your bulk stats by median scene depth - 8K clips show higher absolute error here and I suspect distant scenes rather than resolution.",
  "counts": {
    "4k_valid_at_2p5": 0.632,
    "4k_err_at_2p5": 0.0425,
    "8k_valid_at_1p0": 0.555,
    "8k_err_at_1p0": 0.0809
  }
}