Night
One squirrel, five wires, and a Raspberry Pi learning to tell him from the leaves.
A camera in a window watches a power pole all day. Everything that moves gets a box; a small model on the Pi decides whether the box is him, and every night the Pi retrains that model on what it got wrong. Nothing is streamed live: this page shows what the camera has worked out, and blurred clips of his visits.
01 Caught on camera
His visits, with what the camera saw
Up the pole, and a seat on top
Along the high wire, up the pole, and onto the very top.
02 How the camera sees him
From a moving box to “that’s him”
- 01
Camera
A USB webcam in a window, watching all day.
1280×720 · 30 fps
- 02
Tracker
A box around anything that moves on the pole or the wires.
3,492 things since 16 Oct
- 03
Checker
A small image model on the Pi decides, twice a second: is the box him?
YOLO11n · v4
- 04
Teacher
A bigger vision model on a PC double-checks the Pi.
19,931 crops judged
- 05
Reviews
Aarav, or Claude looking at the frames, settles the close calls the two disagree on.
94 of 94 visits checked
- 06
Retrain
Every night, on the Pi, from everything labeled so far.
9 retrains · back on v4
The tracker keeps a box around everything that moves. The checker is shown a small crop around each box, about 100 pixels across, and scores it.
100%him, on the high wire
1%leaves, rightly ignored
91%a bush it called him
- On the Pi
- Tracking, the checker and the nightly retrain all run on a Raspberry Pi 5.
- A teacher, not a label factory
- A bigger model on a PC double-checks the Pi, and the disagreements come to Aarav, or to Claude.
- Gated
- A new checker only goes live if it beats the old one on crops neither has seen.
03 Today, on the pole and the wires
Where he went
- 9:09 AM24 s
- 10:05 AM41 s
- 10:07 AM23 s
- 10:09 AM27 s
- 10:22 AM26 s
- 10:59 AM11 s
- 11:44 AM49 s
Each line is one visit, drawn from the box the camera kept around him; the latest is the boldest. The shaded bands are where the camera looks for him.
The interest
On top of the pole · 22 Oct, 10:07 AM
From the visits Aarav or Claude checked
- 94visits Aarav or Claude checked, of 94 the camera counted
- 17 Octfirst seen
- 7 min 9 slongest visit, 18 Oct at 8:05 AM, sitting on the pole top
- 7:42 AMearliest he’s come, 18 Oct
- 5:26 PMhis latest bedtime, 19 Oct
- 8:37 AMa usual first visit, 1 h 35 min after sunrise
- 4:08 PMa usual last visit, 1 h 46 min before sunset
- 29visits on his busiest day, 20 Oct
- the polehis favourite spot, 59% of his time
- 6 daysevery day since he was first seen
- 21 san average visit
- 18 min 20 ssitting still, all told
Is it the interest? · 1 of 12
Him, or not?
Real crops, as small as the checker sees them.
You 0 · Checker 0
94 visits in 6 days, and 1 h 14 min of him on camera.
Everything the camera has worked out about him since 16 Oct, counted only while it was really watching. It refreshes every couple of minutes while the Pi is up.
- 94visits of his
- 1 h 14 minon camera
- 6days he came
- 3,492things that moved, each one judged
- 19,931crops the teacher double-checked
- v4the checker running, after a rollback on 22 Oct
01 When he comes
A morning squirrel
Half his visits fall between 8 AM and 11 AM, and he usually turns up 1 h 35 min after sunrise.
One row a day, one column an hour, each cell as dark as his visits in it. The bars on top add up every day.
02 Where he goes
Mostly the pole
- Pole59%27 min 33 s over 67 visits
- High wire20%9 min 15 s over 67 visits
- Top wire12%5 min 29 s over 49 visits
- Middle wire9%4 min 14 s over 16 visits
- Thick cable1%34 s over 8 visits
How he spends it
03 How it learns
A new checker most nights, and one step back
Found him, on crops neither checker had seen
False alarms, on everything else
- v1turned down
- v2turned down
- v3turned down
- v4went live
- v5rolled back
- v6rolled back
- v7turned down
- v8turned down
- v9rolled back
9 retrains so far, and 4 passed the test. Then on 22 Oct v5, v6 and v9 were rolled back: they had learned the bare pole and orange leaves as him, and the test had been built from the same labels. The camera runs v4 again.
- the live checker that night
- the new one, better
- the new one, worse
- within half a point
Each night’s new checker is tested against the live one on the same crops. It goes live only if it finds at least as much of him and has no more false alarms, within half a point, and is better at one of the two. Point at a retrain to see why it went the way it did.
v9 · 22 Oct · rolled back · the latest
Passed its test: it found more of him (83% against 78%), with fewer false alarms (0.5% against 1.7%). Then rolled back: it had been taught on labels that called a whole track him, leaves and bare pole included, and tested on the same.
- Found him
- 78% → 83% better
- False alarms
- 1.7% → 0.5% better
The teacher agrees with the checker on 91% of the last 500 visits it has looked at.
The marketplace
Coming soon
His best moments, caught by the camera, edited and framed as prints.