Blueprint | Depth and reprojection

08Depth-warp study

Depth reprojection measures support beyond each solved camera.

Eight registered depth-warp chains tried step sizes from 0.5° up to a single 30° camera move, each one tracking how much of an output pixel still traces back to the solved source frame.

The corrected measurement is pleasingly blunt: one 30° move keeps … of the frame photograph-derived, while ten 3° steps keep …. Small steps open small holes, and each step then paints over the inventions of the one before it.

The source frame: the house, sharp, from the solved camera

Source reference

The solved frame

invented: 0.0%
camera: r4_0031, solved

Unmodified source frame from the fused solve.

After a single 30 degree warp the house is still legible under heavy speckle

Uncorrected metric

30° in a single warp

invented: 33.5%
largest hole: 16.2% in one piece
steps: 1

Two thirds of the frame remains source-derived.

After ten three-degree warps the house has smeared away entirely

Uncorrected metric

30° in ten 3° warps

invented: 73.0%
largest hole: 1.1%
steps: 10

Each step opens a small region, but the accumulated chain is 73% generated fill.

Re-rendering depth from the splat at every 3° step lifted retained source pixels from 27.0% to 33.4%, which settles where the loss lives: mostly in the compounding of invented fill, only partly in reused depth.

$Nine chains plotted as source-derived pixel fraction against cumulative azimuth. All eight multi-step chains fall below the 50 percent threshold before 30 degrees; the single 30 degree jump ends at 67 percent.$

Every registered multi-step chain crossed the 50% invented-pixel threshold before 30°. Under the original measurement, the single-step arm remained above it. The depth-oracle diagnostic is reported separately.

08bSingle-warp reach

Available reach depends on the source frame and camera direction.

Direct reach was measured by invented-pixel fraction and largest contiguous hole. On r4_0031, at the far end of the arc, the two directions differ by … points: one requires … invented area and the other …. Other source frames reverse or eliminate that asymmetry, so reach is measured per warp. The lateral-dolly control required more generated fill than either orbit direction on every tested frame.

$Six panels: three source frames across, invented pixel fraction on the top row and largest single hole on the bottom, each plotted against azimuth out to 30 degrees. In every panel the dolly control is worst. On r4_0031 the outward orbit costs 33 percent against the inward 25; on l100_0023 the two directions nearly coincide at 32 and 33 percent; on r4_0012, inside the arc, they are 30 and 26.$

The accented curve leaves the solved camera arc and therefore extends coverage. On r4_0031, the inward direction has lower invented area but returns to observed azimuths. Hole topology also differs by direction: … separate holes against …. Both total invented area and contiguous-hole structure affect inpainting.

Superseded reach estimate: the solved frames cover … of azimuth, …. The 30°-limited measurement in this subsection used an incomplete visibility test. Section 08c applies the corrected facing rule and reports the current orbit-coverage estimate.

This study measures source-pixel provenance. Pixels outside the inpaint region remain unchanged, and no reconstruction was trained from these frames. All tests ran locally at $0.00. Data: dwi-chains.json and dwi-reach.json.

Legacy visibility measurement

The original visibility metric reported 66.5% source-derived pixels for one 30° warp and 27.0% for ten 3° warps. Files outside each inpaint region reproduced byte for byte. Section 08c preserves and corrects those values by rejecting back-facing surfaces.

08cCorrection

A facing test stopped back-facing walls from lending pixels they could not have.

The original curve had 120° of orbit scoring … points cheaper than 90°. That impossible reversal exposed a measurement error. Back-facing facade pixels were passing straight through the unobserved building volume. The corrected rule accepts a source pixel only when its surface faces the destination camera, after sky masking, median-filtered depth and larger splat footprints were each ruled out as the cause.

Thirty degrees of orbit, measured without the facing test: the house under heavy speckle

Uncorrected metric

30°, no facing test

invented: …
legacy interpretation: 66.5% source-derived

The uncorrected metric accepts material whose surface faces away from the destination camera.

The same thirty degree orbit with the facing test on: visually almost identical, with slightly more of the frame left empty

Corrected

30°, facing test on

invented: …
visual difference: subtle at 30°

At 30°, the visual difference is small even though the corrected source-derived fraction is lower.

The corrected thirty degree frame with every pixel the facing test withdrew painted in rust red. The withdrawn set follows the right-hand wing of the house and the roof edge; the porch face, still pointed at the camera, comes through untouched — Rust marks pixels removed by the facing test, totaling … of the frame. Rejection follows the receding wing and roof edge while preserving the camera-facing porch. Foliage accounts for … inside the densest 30% window, so unstable leaf normals contribute some noise; the building surfaces show the same directional pattern.

A hundred and eighty degrees of orbit, no facing test: the camera stands behind the house and is shown the front facade through the walls

Failure case

180°, no facing test

invented: …
invalid content: front facade visible from the rear

Without facing rejection, the 180° rear view receives a lower invented-pixel estimate than the 90° side view.

The same hundred and eighty degree orbit with the facing test on: almost the whole frame is empty because the front-only capture contains almost no rear-elevation evidence

Corrected

180°, facing test on

invented: …
known source coverage: almost none

A front-only capture provides almost no valid surface evidence for the rear elevation.

$Invented pixel fraction against orbit angle out to 180 degrees for two source frames. The dashed uncorrected curve rises, turns back down past 90 degrees and on r4_0031 never crosses half the frame invented. The solid corrected curve rises throughout and crosses at 57 and 44 degrees.$

The uncorrected curves decline beyond 90° because they accept back-facing pixels. With facing rejection, … of four measured arms rise monotonically, compared with … previously. Corrected values are lower-bound estimates because normals derived from splat-rendered depth remain noisy.

The corrected facing test lowers retained-source estimates for both decisive arms. The original single-jump value is …; the original ten-step value is …. The gap increases from … to … percentage points, strengthening the conclusion that repeated warping loses source evidence faster. Legacy plates retain their original labels for reproducibility.

Plan view of a full orbit of the property. Forty degrees is solved, a hundred and one more degrees are reachable by one outward warp from each end frame, and the remaining two hundred and eighteen degrees have no solved frame that can reach them at any angle. — The reachable band combines the solved arc with one outward warp from each end: … past r4_0031 and … past l100_0023, cutting off at half the frame invented. That comes to … of azimuth. The remaining …, … of a full turn around the property, has no source frame within measured reach. The limit follows the recorded camera positions.

Under the corrected facing test, 60% of the orbit has no reachable solved source frame. Depth warping genuinely extends the observed front arc, and the missing side and rear elevations need new evidence from somewhere else entirely. This local study cost $0.00. Data: dwi-far-reach.json.

Facing-test provenance

The corrected figures derive from the sealed r3 manifest. Payload hashes for the four comparison plates are recorded in dwi-facing-plates.json.