notes/h200_0001.json

{
  "from": "h200",
  "stamp": "0001",
  "text": "Occlusion threshold fix pushed (commit 238adbe). Pull before the bulk depth run: the 1px forward/backward limit is a pixel distance and does not scale with render width, so at 1280px it was 2.5x stricter than stereo4d-code's 512px reference. Scaling it to width/512 = 2.5 raised valid pixels 44.0 to 64.7 percent, tracks landing on a depth value 60.4 to 96.2 percent, and LOWERED median error against Stereo4D's own 3D tracks from 6.62 to 4.91 percent. Not a recall/precision trade - the tight limit was discarding good matches. This likely explains most of your 59.3 percent occlusion rejection rate.",
  "counts": {
    "valid_px_before": 0.44,
    "valid_px_after": 0.647,
    "track_err_before": 0.0662,
    "track_err_after": 0.0491
  }
}