The patio at night through the glass door: string lights over a stone wall, a plate of food on the pavers, and a raccoon mid-stride along the top of the wall.

Backyard Critter Cam · Field reportJun 7 – 26, 2026

Twenty
Nights

One camera at a sliding glass door.
Everything that moved, measured.

photographs
32,214
visits
652
species named
30
raccoons known by name
5

01 · The distillation

The camera counts photographs.
The journal counts visits.

A lingering raccoon can fill a memory card without telling you anything new — the season's record is 707 photographs of a single visit, Notch on the first night, June 7 at 10:15 pm. So consecutive sightings collapse into visits, and twenty nights of shutter become a number a person can actually read.

32,214
photographs saved
652
visits, once bursts collapse
5
raccoons known by name

Alongside: 2,997 behaviour clips, 30 species named, 11,077 appearance vectors — all from one webcam and one SQLite file. data/meta.json

02 · The night shiftn = 118 raccoon visits

The yard clocks in at dusk.

Of the 118 raccoon visits with an arrival time, 98% arrive between 7 pm and 7 am. The rush peaks in the nine-o'clock hour; a typical stay runs five minutes. The busy window wraps around midnight — so the chart does too.

Arrivals, around the clockevery raccoon visit

Bar length is the count of visits that began in that hour; the shaded arc marks 7 pm – 7 am. The two daylight arrivals are ringed — the kind of visit the dashboard flags as off-pattern. data/twoaxis.json

See the numbers
22
arrivals in the peak hour — 9 to 10 pm
5min
median stay, door to door
2am
Notch's peak hour — the late shift is his

03 · The cast5,136 named photographs

Five names.
One very unequal guest list.

Human confirmations put a name on 5,136 photographs — and Stan and Notch account for nine in every ten of them. Pepsi appears forty-nine times less often than Stan. Fame in this yard is a long tail.

Photographs per named raccoonconfirmed only

Portraits are real crops from the export; Miss B. and Pepsi were photographed too rarely to make its public slice. data/meta.json

04 · Appearance5,356 pairwise comparisons

The camera remembers the night better than the animal.

Compare every pair of the 104 raccoon photographs in this export, in the re-ID model's own similarity space. Frames from the same visit look alike — 0.51 on average, with bursts touching 0.99. The same raccoon on a different night? 0.23 — barely ahead of two different raccoons at 0.15. Pose and backdrop outweigh identity, and a single photograph cannot be matched across nights.

How alike two photographs lookcosine similarity, 0 → 1

Each row is the distribution of one kind of pair; darker means more pairs land there. Ticks mark the mean. data/appearance.json

See the numbers
The same test, on visit prototypes16 visits, 120 pairs

Separated, but overlapping — which is why the system suggests and a human confirms. data/appearance.json

See the numbers

05 · The gate320 scored crops

A species model can't say “nothing.”

BioCLIP names species — show it an empty deck or a plate of kibble and it must still pick one. (For a while, every false alarm was a “brown rat.”) So a general CLIP goes first and answers the only question that matters: is this an animal at all? On 320 scored crops the two populations never touch: every real animal scores below 0.43, every non-animal above 0.60. The threshold sits in open water — zero crossings, in either direction.

“Probability this isn't an animal,” per cropgate model, 0 → 1

One dot per crop. The hairline at 0.60 is the live rig's actual cut. data/decoy.json

See the numbers

06 · Gait7 measured walks

Every walk keeps time.

From each behaviour clip, the tracker reads the body-bob of a walking raccoon and recovers a stride — from a 0.8 Hz amble to a 3.3 Hz trot. Path straightness spans just as wide: 0.13 (milling about the plate) to 0.88 (a beeline across the wall). A limp would read the same from any camera angle — which is what makes motion the second axis of identity.

Stride cadence, walk by walkhover a walk — its dot keeps its beat

Each dot is one tracked walk; the pulse repeats at that walk's measured cadence. data/gait.json

See the numbers

07 · Confidence, auditedgraded against human verdicts

Believe the label by day.
Squint at it by night.

Grade the model's confident species labels (≥ 0.8) against every human verdict, and daylight keeps its promise — night doesn't. Through the glass, infrared is monochrome and soft, and a confident label in the dark is far likelier to be wrong than a confident label at noon. The camera didn't reveal that; the audit did.

96%
labels right in daylight
63%
labels right after dark

Measured July 1 on 165 gradable rows — 138 day, 27 night, 129 of them raccoon. One camera, three weeks, mostly one species: read it as the shape of the problem, not a benchmark. The eval reruns in one command (python eval.py) as the corpus grows.

The season is still being written.

Every number on this page is computed live from the same frozen export that powers the making-of demos — regenerate the export and this page follows. Behind the glass, the journal keeps going.

Behind the Glass — how it all works Source on GitHub

Provenance

  • Toplines & cast — data/meta.json
  • Arrival clock, stays — data/twoaxis.json (118 raccoon visits)
  • Similarity distributions — data/appearance.json (104 crops · 5,356 pairs · 16 prototypes)
  • Gate populations — data/decoy.json (320 crops scored with the live gate model)
  • Stride cadences — data/gait.json (7 tracklets from behaviour clips)
  • Record visit & un-blend — data/reid_game.json
  • Day/night accuracy (Jul 1) and re-ID ROC-AUC (Jul 18) — dated eval.py runs; reports/ is deliberately gitignored, so reproduce them on your own corpus rather than quoting these.