Instrumented People, Places and Things: the napkin dispenser study, a model of the shop

Cameras first. A sparse rebuild (COLMAP) from the 85 originals, with the lens's calibration held fixed, solves the main room and the washroom as separate pieces. A looser run that claimed 84 photographs in one piece had joined the washroom to the main room on false matches, and was dropped. The rebuild is evidence for where the cameras stood; it is not the model.

The room's directions come from straight lines in the photographs: 8,227 long edges give one vertical and two wall directions, square to each other within 0.4°.

The window bar's photographs form a piece of their own, joined to the room at one pair of photographs taken 9 s apart. The turn comes from their matches, the size from the floor tile, the place from the left wall's plane. Then the tile grid is the test: the bar piece's grout lines fall within about 1 in of the room's grid carried past where it was read.

A grid repeats, so it can mislead. DSC_0068 was solved from tile crossings and fit as well one tile off, which put the bar's end 13 in too far along. The far views disagreed. One tile over, the bar's end and the sill's corner agree with DSC_0071 within about an inch.

The model is read by hand: a corner or edge is read in two or more photographs, or dropped onto a plane already known, and written down in inches. Nothing is fitted to the cloud of points. Where a read is an estimate, its card says so.

Solved from 14 fixed points read by hand in a median of 60 frames, which takes out most of the people: ceiling lamps, the bulkhead's corner, the counter, the freezer's panel, a floor vent. They fit to 1.8 px at 320 wide. Four corners of a menu board, held out, land 0.2 to 3.5 px from where they are read, up to about 4.6 in on that wall. Leaving out any one anchor moves its place by at most 2.3 in.

The focal length and the distance along the view trade against each other. The best fit alone puts the camera about 8 in outside the glass, which cannot be, so its place is held inside: x 22, y 1, z 54 in, and the anchors allow about y 1 to 9 along the view. Its field of view is about 60 by 47°. That is about 10 in above the bar's top and against the glass. No photograph shows it there, so its being on or at the bar is inferred.

The frame itself does not move through 22 March: eight stretches of the day agree to 0.13 px. The frames stay off this page, because they show customers.

Corners read by hand in photographs that gave the model no read of them, against where the model puts them. A photograph with a person in it is checked here and shown nowhere.

Reading the model over the photographs also corrected it. Holders D and E were first read from DSC_0016, placed by its turn from DSC_0015, which leaves its place along the bar free: the overlay put holder A where DSC_0016 shows none. They are now read from DSC_0068 and DSC_0094.

An automatic search for the nearest edge in the picture could not tell the model from a copy of it moved 60 px, because the room is so cluttered, so it is not used.

The photographs are from 3 February and the webcam's frames from 22 March. Seen through the webcam's solved place:

• In place, to within what 320 by 240 can show: the counter, the bulkhead, the freezer's panel, the floor vent, the ceiling lamps, the menu board.
• Holder F's spot: an object of that size stands there on 22 March, bluish at this resolution. It cannot be called a holder from these frames.
• Changed: the paper hearts over the counter are other hanging decorations by March, in about the same spots. Close to the webcam, at the frame's lower right, stand a cyan and pink sheet and a yellow card that the February photographs do not show.
• Out of the frame: four of the five holders on the bar. Holder C's February spot falls at the frame's right edge or just past it, within about a foot of the webcam, so it is too close to call. Only a corner of the bar's front edge enters the picture, so these frames cannot say whether a holder moved.
• Not compared: tables and chairs, which are not in the model.

• The modules. The study's readings say that 102 and 104 shared a surface or object, and little else about where any sat. Their candidates are lit as a reading drawn here.
• The washroom against the shop. No photograph sees both, so it is drawn to one side, and its place there means nothing.
• Six photographs have a direction and only a borrowed place, because they show blank wall, ceiling or sky: DSC_0017, 0018, 0023, 0085, 0086, 0087. DSC_0069 is tied by a long chain, good to about 7 in.
• The logging PC. No photograph shows it, since the photographs come before it.