Casa — property capture

84 & 90 Spruce St · objects · the bank. Works with no signal once loaded.

Checking offline copy…

The one rule that ruins takes

Lock exposure AND focus before every take. Press and hold on the subject until AE/AF LOCK appears. Autofocus hunting mid-clip changes the effective focal length, and the solve assumes one camera with one set of intrinsics for the whole clip. Auto-exposure riding up and down as you swing past a bright wall makes one surface look like two.

If you remember nothing else on this page, remember this one.

Before you leave the house

Set these tonight, not in the driveway.

Which iPhone do you have? — 30 seconds

Open Settings → General → About → Model Name.

Any “Pro” or “Pro Max”, 12 or newer → you have LiDAR. Everything below still applies, and a LiDAR capture app can additionally record depth and camera poses directly, which is a materially better input than video alone. Video is still the reliable path — shoot it either way.
Anything else → no LiDAR, pure photogrammetry. Same protocol, and lean harder on the loop closure and the extra height passes, because there is no depth prior to fall back on.

Do not delay the shoot over this. The video protocol below works on both.

Three different jobs — do not mix them

What you are pointing at	How	What comes out
A structure — house, shed, trailer, shipping container	One continuous orbiting video VIDEO	Gaussian splat — a walkthrough and a visual record
A small object — a fixture, a valve, a broken part	One still photograph STILL	A mesh — watertight geometry you can design and print from
Terrain — the bank behind the house	Parallel video sweeps TERRAIN	Surface topography — slope, break lines, cut volume

A splat is not a mesh. Splats are excellent for walking through a space and poor for design, which needs watertight geometry. Optimising for one starves the other — one reconstructor takes a single photograph and infers a whole object; the other takes a hundred views and infers where the camera was. That is why they are separate takes with separate names.

Scale — the thing that cannot be added later

A reconstruction has no idea how big anything is.

The same solve fits a doll's house and a cathedral. Without a known length in the frames, you get a shape, not a measurement — no slope angle, no cut volume, no foundation dimension, nothing you can plan a dig or a build against.

So: a known-length object in shot for at least 5 seconds, from more than one angle — and its length written down against the clip name. A tape measure extended and locked, an 8 ft 2×4, a folding rule. Say the length out loud while filming so it is on the audio track too.

Thirty seconds per take. It cannot be recovered afterwards.

Your shipping container is a free scale reference for the whole site — standard containers are 20 ft or 40 ft externally. Measure yours once, write it down, and it anchors everything shot near it.

How to walk it VIDEO

Setting	Value	Why
Resolution / rate	See the ladder below	Higher frame rate shortens the exposure per frame, which halves the motion blur. That is the reason — not the frame count.
Lens	Main 1× rear only	0.5× ultrawide barrel distortion does not fit the camera model. Never zoom — the solve fits one set of intrinsics to the whole clip.
Orientation	Landscape	Wider field per frame.
Walking speed	~1 m/s outdoors · 0.5 m/s indoors and around corners	A deliberate stroll, about one pace a second. Slower where the solve struggles. Slowing down does not rescue dim light — see below.
Take length	60–90 s for an exterior orbit. As long as it takes indoors — often 3–6 min.	One structure, one unbroken take. The 60–90 s figure is for orbiting one building; do not try to force a whole floor into it — see below.
Grip	Two hands, elbows into your ribs	Whipping the phone causes rolling-shutter shear.

Which mode — the ladder

Pick the highest one your phone offers. Sharp beats big: a blurry 4K frame is worth less to the solver than a sharp 1080p one, because it is matching features, not admiring resolution.

4K 60 — best. Most detail and the shortest exposure per frame.
1080p 60 — better than 4K30 while you are moving. Half the blur beats double the pixels.
4K 30 — fine, but walk at half speed to get back the sharpness you gave up. Roughly 0.5 m/s outdoors.
4K 24 — do not use it. 24 fps is a cinema frame rate, chosen because it renders motion blur pleasantly. Motion blur is precisely what destroys a reconstruction.

Check under Settings → Camera → Record Video. If 60 fps is there at any resolution, use it.

Overlap — ignore the percentages

Published overlap figures disagree — some guides say 20–30 %, others 70–80 %. Both are written for stills. At 60 fps and walking pace, consecutive frames already overlap by well over 90 %, so a percentage is not a useful target.

The real target: three or more genuinely different angles on every surface. That is what the second and third height passes buy you, and it is what you should be checking as you walk.

The moves

Close the loop. Walk the full circuit and keep filming 3–4 seconds past your starting point. Closure needs the same position and a return viewing angle within about 40° of the original — position alone is not closure.
Keep walking rather than turning on the spot. Pivoting in place gives the solver rotation with no parallax, which is nearly worthless.
Height passes. One at chest height, one low (waist or knee, catching the base of the walls and the ground), one high with the phone above your head if the eaves matter. Same loop, different height. Change height gradually, over 3–5 m of travel.
Ground plane in shot throughout. Not walls only — the ground is what ties the passes together.
Do not skip the boring sides. A blank wall you skipped is a hole in the model.
Never stop to “get detail.” A static camera adds no new viewpoint, and duplicate frames degrade sequential matching.
Doorways are the highest-failure transition. Slow from 2–3 m before, resume 2–3 m after, and leave the door half open. If the doorway is tight, turn your shoulders sideways to fit through — keep the camera pointing the way you are walking. Rotating the phone 90° mid-take shears the frame and breaks the overlap between consecutive frames, which is the one thing the whole take depends on.
Indoors: one continuous take for the whole floor, not one per room. Walk into a room, do its perimeter along the walls, then a pass through the centre, then let the camera rise and fall to catch floor and ceiling — then walk out through the door and into the next room without stopping the recording.
Expect an interior take to run 3–6 minutes, and that is correct. The 60–90 s figure is for orbiting a single building from outside. At 0.5 m/s, 90 seconds is only 45 m of walking — nowhere near a floor of a house with three passes a room. If you try to force a floor into 90 s you will either sprint, which blurs every frame, or stop recording halfway, which breaks the take in two.
Multi-storey: start at the bottom and work up. Walk the stair run while raising and lowering the phone — do not tilt it from the landing.

Light

Overcast or open shade is the best condition there is for this — not a compromise. Otherwise early morning or late afternoon.
Avoid the hard sun/shade boundary. The reconstruction bakes the light into the surface, so a shadow edge becomes permanent geometry-coloured wrongness.
Blown windows and white sky give the solver nothing.
If the sun comes out mid-orbit, redo the pass. Consistency across one take beats absolute quality.
Indoors: all lights on, blinds open, but never point the camera at a window or a lamp.
Do not film in low light at all. The phone answers dim light by holding the shutter open longer, and every frame smears. Walking slower does not fix it — the exposure is what is wrong, not your speed. Turn lights on, open a door, bring a work light, or come back when it is brighter.

Blur is checked automatically, and a bad take gets rejected

Frames are scored for sharpness and the smeared ones are dropped. If more than about a quarter of a clip is rejected, that clip needs re-shooting — slower, or in better light.

You will not find that out until the footage is processed, which is the whole reason the throwaway test comes first. Get one clip through the pipeline early and you learn whether your pace and your light are working while the day is still ahead of you.

What will not reconstruct

Shoot it anyway — just expect holes, and do not waste the day fighting it.

Mirrors, glass, windows, polished floors, chrome. Stay 1 m back from reflective surfaces.
Featureless white walls. Put something textured on them — painter's tape, a towel, anything — or they will not solve.
Anything that moves: people, dogs, curtains, washing on a line, a tree in wind. It comes back as drifting fog.

Redo the take if…

The sun came out, or went in, partway round.
You saw the image pulse or breathe — that is autofocus hunting. Re-lock and start again.
You stopped recording partway through the structure.
Someone or something walked through the shot.
You pivoted on the spot instead of walking an arc.
Playback shows blur, blown windows, or a side you missed.

A redo costs 90 seconds. A bad take costs the trip back.

The bank behind 84 TERRAIN

You cannot orbit a hillside, so this is its own technique.

Still ONE continuous recording — do not stop between sweeps.

It is tempting to shoot the near pass, stop, walk back, and shoot a second. Do not. The processing assumes one unbroken path in capture order; separate files either overwrite each other or make the solver teleport between the end of one sweep and the start of the next, which it cannot recover from.

Walk it as one take: out along the near line, then — still recording — turn and come back along a line further out, overlapping where you started. Then out again further still if the face is big. A there-and-back with an offset is a loop. A one-way pass is not.

2–3 passes at different distances from the face, all inside that single take.
Capture the top of the bank and the toe of the slope explicitly. Those break lines define the cut, and they are what people miss.
Overlap where you turn around and film a few seconds past it — that overlap is what closes the loop and stops the hillside warping.
Scale reference sized to the pass. A tape measure is unreadable from the far line — use something big out there: the container, a vehicle, an 8 ft board stood up.
Vegetation is the enemy — bare ground reconstructs, leaf cover does not. If this matters, late autumn or early spring is materially better than summer.

What this will and will not tell you — read this before you plan anything

What the pipeline actually produces is a Gaussian splat. Once it has a known length in it, you can measure distances and read the slope off it, and you get a faithful visual record of the bank as it stands today. That is genuinely useful and it is the honest claim.

It is not a volume. A cut-and-fill figure needs watertight surface geometry, and a splat is not that — extracting a proper surface from it is a further step that is not built yet. An earlier version of this page promised you a volume. That was wrong, and this corrects it.

And it tells you nothing about soil composition, bedrock depth, or the water table — which are what actually decide whether anything can be built into that bank. Test pits and a geotechnical engineer answer that, plus a New York permitting question on top. Do not let a good-looking model imply the dig is feasible.

Small objects STILL

Completely different discipline. The object reconstructor takes one photograph — orbiting, bracketing or filming produces nothing it can use.

One object, filling roughly 70–80 % of the shorter side.
Plain, matte, mid-tone background. Bare bench timber is ideal. Nothing patterned, shiny, or the same colour as the object.
Flat, even light — open shade or overcast, or indoors away from a window. No flash. Hard sun makes a black shadow that reconstructs as a hole.
Camera level with the object's middle, straight on. Not looking down at it from standing.
Nothing else in frame — especially not your hand. Stand it against a block or lay it down and step back. If it is in the photograph, it will be in the model.
Shoot three or four and pick the best. Do not stitch, orbit or bracket.

Grease and wear are fine, and arguably useful. Loose grit and hanging rags are not — they become geometry.

The order to shoot in

Light changes through the day and exteriors care about it while interiors do not. That single fact sets the running order — spend the good light outdoors and the harsh light indoors.

When	Shoot	Why then
First thing (before anything else)	The throwaway test — one shed or the container, 90 s, full protocol	So the pipeline gets validated while you keep working. If something is broken, this is when we want to know — not tonight.
Early morning soft, low sun	Exteriors — 84, then 90. All three height passes each.	The best exterior light of the day. Do not spend it indoors.
Midday harsh, high sun	Interiors, room by room. Objects in open shade.	The one part of the day when harsh sun genuinely does not matter.
Late afternoon soft again	The bank, then any exterior you missed.	The bank wants even light with no hard shadow raking across the slope — a shadow edge becomes permanent false geometry.

If the whole day is overcast, ignore this table and shoot in whatever order suits you. Flat overcast is the best condition there is for every one of these jobs, and it removes the constraint the running order exists to work around.

When you get back

Do not delete anything from the phone until it is confirmed off it. Not after the copy appears to finish — after it is confirmed.

Get the clips onto the box by whichever route is easiest that day — the laptop over Tailscale, a Drive folder, or straight upload. All three land in the same place.
GPS is stripped on arrival, before anything else touches the files. Your address is in that metadata and it is not going anywhere public.
Frames get pulled at the sharpest moment per time-bucket, blurred stretches dropped, and the filenames stay in capture order — that ordering is load-bearing for the solve, so nothing renames them.
Then reconstruction, one scene at a time, cheapest first.

Keep the original clips. Everything downstream can be re-run; the footage cannot be re-shot.

Takes

Tick as you go. Saved on this phone, no signal needed. Add takes for anything not listed.

Name the clips as you go

In Photos, or on a scrap of paper — either is fine, but do it the same day. Frames land in per-property folders and the name is how they get there:

84spruce-ext-chest · 84spruce-kitchen · bank-sweep-near · obj-container

And next to each one, the scale object and its length. bank-sweep-near — 8 ft 2×4.

Honest status of what happens next

Frame extraction from your video is built and measured — it picks the sharpest frame per time bucket and rejects smeared ones, graded against known-good and known-blurred test footage.

The GPU reconstruction has never been run. Not once. Volume mounting is the highest-risk unknown, and the blur threshold is calibrated only on synthetic blur, not real handheld motion blur.

So shoot one throwaway structure first — the shed, or the container — and let that one go through the pipeline before you spend the day on the houses. If something is wrong, that is when we want to find out.

Property media stays on-box and is GPS-stripped on arrival. Nothing with your address or interior goes to a public service.