TABOR

"Tabor" is Polish for rolling stock. Warsaw runs about 2 600 buses and trams in regular service, split across 44 models. Some models have hundreds of vehicles, and some have four. On top of that come 36 kinds of museum and club vehicles, Tramwaje Warszawskie's works cars, and the odd bus on trial. Every vehicle carries a fleet number, and ZTM's database says which model each number belongs to. So a photo of a bus is enough to know exactly what you've seen, and a list of every number is enough to know what you haven't.

TABOR does both halves. CATCH reads the fleet number off your photo, on the phone. BOOK files the vehicle under its model and batch. HUNT shows every vehicle you haven't caught yet, live, on a map, and follows one to your stop on the Lock Screen. ME keeps the score.

---

Install

Get it on the App Store: apps.apple.com/app/tabor/id6816408610 (iPhone, iOS 18 or later, free). Bug reports and ideas go in issues, and questions in discussions. CONTRIBUTING.md says what makes a report or a pull request easy to act on.

Po polsku: opis i instalacja na thefilip.com/tabor.

Or build it yourself. Open the project in Xcode 27 and run it on your phone:

open Tabor.xcodeproj

Set your own team under Signing & Capabilities for the Tabor, TaborControls and TaborShare targets.

The live map and the live-assisted number reading use Warsaw's open data. The app doesn't carry a key: it asks TABOR's proxy (proxy/, a Cloudflare Worker running in Warsaw), which holds the key, keeps each copy of the feed for 3 seconds so phones share one call to the city, gzips it, and allows each address 60 calls a minute. Builds from this repo use that proxy as they are. To go direct instead, paste a free key from dane.um.warszawa.pl into Settings › Live data. To run your own proxy, change the domain in proxy/wrangler.toml and TABOR_LIVE_PROXY in Config/Tabor.xcconfig, then:

cd proxy && npx wrangler secret put DANE_TOKEN && npx wrangler deploy

---

CATCH

• The number is read from the shot. Take the photo and Vision's text recogniser reads the part inside the brackets, on the phone. If the whole frame gives nothing, or only a short number, it's read again as nine overlapping tiles, so a small number on a whole-vehicle shot has enough pixels. The live preview isn't read at all: that only cost battery and heat, and you can see the bus coming anyway.
• Only fleet numbers count. A photo of a bus is full of other digits: line numbers, times on the destination display, and the registration plate.
  • Runs glued to letters are dropped, and so are runs after a two- or three-letter county code (WX 2043F, WPI 12345), and plate digits that start with 0. "Nr 1974" still counts, and so does a number with paint glued on.
  • Times and decimals (20:15, 3.14) are dropped.
  • A number the database doesn't know is only kept if it's four digits long.
  • One letter that looks like a digit (1O23, I974) is read as that digit, but only when that makes a number in the fleet, and it ranks below a clean read.
  • When a photo shows only a short number, it's read again more closely, so a line number isn't taken for the fleet number.
• It knows what's around you. With a location fix and the live feed, a number that belongs to a vehicle within 300 m gets a boost. If the read is one digit off a vehicle within 150 m (OCR turning 4235 into 1235), that vehicle is offered too, though a clean read of a real number still wins. When a number belongs to both a bus and a tram, whichever one is actually nearby settles it (otherwise the reveal asks which one it is), and the catch gets its line number from the feed.
• More than one in the shot. Other vehicles in the photo are offered on the reveal as chips ("+ 5918 · ALSO IN SHOT"), each with its own sticker, and saved as catches of their own. With the live feed they have to be running nearby.
• Coupled trams count twice. Konstal 105Na, 105N2k and Cegielski 123N run as two cars, each with its own number, but the live feed reports a set under one of them. So the other car isn't "fixed" into its neighbour: it's kept as read, and it takes the set's line. The reveal offers + SECOND CAR, suggesting the other number (from the photo, the feed or the next number along), and both cars go in the book.
• BUS / TRAM / AUTO is a preference, not a filter. A bus number read in TRAM mode still matches the bus. Spotters forget to switch back.
• The shot is what you framed. The viewfinder is a 3:2 photo frame, and the saved shot is cropped to it. Zoom has one button per real lens (0.5×, 1×, 2×, 5× on a 16 Pro), so 5× is the telephoto rather than a crop of the main lens, and the phone clicks as you cross from one lens to the next. The last button goes to twice the longest lens (10× on a 16 Pro), for a bus that's already pulling away.

The sticker

The photo develops while the vehicle is lifted off the background. That's the same subject-lifting model as the Photos app, running on the phone. The sticker is cut from the object the fleet number is painted on, so another bus or a car alongside doesn't steal it. People standing in front of the vehicle are erased, bike included, unless erasing them would bite into the vehicle. WRONG CUTOUT? on the reveal tries the next object in the photo.

It gets a thick white border that follows its outline, and the sticker slams down with a haptic pattern that gets longer the rarer the model is. A COMMON lands with one thunk. A LEGENDARY builds up in step with its glow, hits on the frame you see it land, and ends in sparks.

A line underneath says what the catch means for your book:

One more — 1987 — and the 2024 batch is done.

Hold the sticker to peel it off the backing and stick it in the book.

---

HUNT

Every bus and tram you haven't caught, from Warsaw's live GPS feed. The feed is read every 10 seconds while the map is on screen, every 30 seconds elsewhere in the app (so trails are ready when you open the map), and never in the background.

• Two kinds of marker. A vehicle on its own is a tag with its line number. Vehicles that would overlap become one bubble with a count. Filled means a model you don't have yet, and outlined means a model you have but not this vehicle. The colour is always the tier.
• Groups hold still. The bubbles are fixed squares on the map, not clusters around the vehicles, so they don't jump about with every update as the buses move. They only regroup when you really zoom.
• Which way it's going. The feed only says where a vehicle is. So TABOR keeps the last ten minutes of positions, laid along the street, and draws an arrow once a vehicle has gone 30 m. Each card says how old its position is.
• Where it goes next. Every line's route comes from ZTM's timetable data (GTFS), rebuilt daily. A picked vehicle is placed on its route: a dotted line shows where it goes next, the card lists its next stops and where its run ends, and the pin creeps along to where it should be now between updates. It's only COMING YOUR WAY if its route actually passes you; otherwise it TURNS OFF before you.
• TRACK it. While a vehicle is coming, TRACK puts it on the Lock Screen and in the Dynamic Island: how far away it is, how many stops are left, then HERE with a button that opens the camera. The proxy follows the vehicle and pushes each update, so it keeps going with the app closed.
• The map follows the one you pick. Move the map yourself and it stays where you put it, and FOLLOW brings it back. With a heading, your dot shows which way you're facing, and the map can turn to match.
• A wanted list. The filter takes buses or trams, whole tiers, particular models, and specs: length, drive and low floor. Any filter but the type alone makes HUNT search the whole city, not just the 3 km around you, and sort by distance: if you're after one particular model, how far away it is matters most. From outside Warsaw, a filter is the way in.
• Search. A fleet number finds that vehicle, running or not. A line lists every vehicle on it right now.
• Three views. UNCAUGHT shows everything not in your book, MISSING only models you don't have yet, and ALL shows everything running, caught or not. A vehicle you've caught gets the same marker as any other model you have, and its card says when it went into your book.
• Rarest first. Unfiltered, the list shows models you don't have yet first, then goes by rarity, then by distance. Vintage and test vehicles sit at the end of the book, but one that's actually out running is as rare a sight as a LEGENDARY, so HUNT ranks them with the legendaries.

---

BOOK

Every model, with every vehicle ZTM lists, split into batches by operator, build year and depot:

2024 BATCH · R-1 WORONICZA · 1970—1987

Owned stickers come first, then the gaps. A batch turns green once you have the whole batch. Each model page has its length, drive, floor, air-con and how many people it carries, every drive it comes in, and SHOW ON MAP for the ones running now.

Tier	What it takes
LEGENDARY	A model with 12 vehicles or fewer
GOLD	48 or fewer
RARE	80 or fewer
COMMON	Everything bigger
VINTAGE	Tourist and museum vehicles, only out on summer weekends and at events
ON TEST	Trial vehicles on loan to an operator, here for a few weeks
WORKS	Tramwaje Warszawskie's own works cars: measurement, transport and welding cars, out on no timetable

Rarity is fleet size and nothing else. There's no guessing at how often a model runs. VINTAGE, ON TEST and WORKS are tiers of their own, whatever their size, and none counts toward the fleet percentage or the "complete a tier" badges. They aren't rare, they're seasonal, passing through, or not for passengers.

A vehicle's page has its age, when you first saw it, how many times you've seen it, any special livery, and every sighting: when, where (district and street), and which line it was on. A vehicle caught more than once shows the picture you pick. Swipe a sighting to fix its number, model or line, or delete it. A deleted sighting can be brought back with UNDO for a few seconds.

---

ME

• The trophy shelf: your three rarest stickers. Then caught, the fleet share (to one decimal place under 10%, so the first few hundred catches visibly move it), the day streak, and how many depots you've caught vehicles from.
• 53 badges, drawn as struck coins. Drag one to spin it: it has a rim and an engraved back, and the phone clicks on every quarter turn. Some are tiered, from bronze up to platinum (Collector, Fleet share, Line hopper, Photographer). Others are for one thing: Unicorn (the only vehicle of its model in Warsaw), Twins (back-to-back fleet numbers), Double life (a bus and a tram with the same number), Dressed up (a special livery), Rainbow day, Every district, Every operator, one badge per depot. Tap one to see the vehicles that earned it, or what's still to find. A few stay secret until you earn them.
• Weather badges (snow, rain, heatwave, deep freeze, a thunderstorm) look up each geotagged catch in Open-Meteo's archive afterwards. That's the one thing TABOR sends off the phone: a catch's time and its rough position (to about 1 km). It can be switched off in Settings.
• ALL STATS: any month, year or all time, with catches, vehicles, models, lines and depots, buses against trams, a calendar of the days you were out, your hour and your day, and your top models, places, depots and lines. It shares as a story-sized card.
• The catch log: every catch, newest first, saying which ones were a new model or a new vehicle.
• Where you spot: a map of every geotagged catch. Tap a pin for that catch's sticker, when and where it was, and a way to its page.
• Share a catch as a 4:5 card: the sticker under a spotlight in its tier's colour, the number tag, and where and on which line you caught it.

Outside the app

• Widgets: Stats (a month, a year or all time, picked when you add it), Memories (what you caught a week, a month or a year ago today), Your book, Latest catches, Rarity sets and Sticker shuffle. Tap a sticker to open that vehicle.
• Lock Screen tracking in the Dynamic Island and on the Lock Screen (see HUNT).
• Catch and Hunt controls for Control Center, the Lock Screen and the Action button. "Catch a vehicle" is also in Siri, Spotlight and Shortcuts.

---

Fleet data

Tabor/Resources/fleet.json is built by scripts/fetch_fleet.py from two sources: the ZTM vehicle database, and the city's open-data vehicle list, which adds numbers ZTM hasn't listed yet and each model's specs. On top of that come hand-kept lists for what neither has yet or has wrong: new deliveries already running, buses on trial, club-owned vintage buses, Tramwaje Warszawskie's works cars, retired vehicles still listed, vehicles filed under the wrong model, coupled trams, and special liveries (confirmed from photos, since the city's paint data misses repaints). The script's header explains each list.

python3 scripts/fetch_fleet.py --offline

Rebuilds from the saved files in data/ in seconds. Leave out --offline to fetch the city's list again, and add --scrape to scrape ZTM too, which takes a while. ZTM blocks GitHub's runners, so the scrape runs from a home connection, not in CI.

Every night a GitHub Action fetches the city's list, rebuilds, runs the tests and commits the result to main by itself. It asks for a human only when more than 20 vehicles would leave at once (a pull request) or the city lists a type no model has yet (an issue). Corrections go into the script's lists, never into fleet.json by hand: a check fails any commit whose fleet.json isn't what the script builds, so the nightly rebuild can't undo a fix.

New buses don't need an app update. The app checks this repo's main for a newer fleet.json once a day and uses it from the next launch. A downloaded copy only wins while it's newer than the one built into the app, so an app update with fresher data is never overridden by an old download. Settings can turn the check off. When a model is split in two, catches move to the right one by themselves.

Hand-added numbers come from a source (Warszawikia, TransInfo, the KMKM club, phototrans.eu) or were seen live on the feed. None are filled in from an order size. As soon as ZTM lists a number itself, ZTM's row wins and the hand-added one drops out. Every source is credited at the bottom of the book.

[!NOTE] The live feed and the database don't always agree. The feed keeps some vehicles' last positions for years, so anything older than three minutes is ignored. A vehicle the database doesn't know never shows on HUNT, because TABOR couldn't say which model it is.

---

Your book

It lives on the phone and nowhere else: there's no account, nothing of yours goes to a server, and no iCloud sync for now. What the app does send (the live feed, the weather, Lock Screen tracking) is listed in the privacy policy. Settings › Backup exports the whole book as a ZIP: every sighting, photo, sticker and hand-picked model. Importing a backup only adds what's missing, so importing the same file twice changes nothing.

The book keeps its own copy of each catch's photo at 1600 px, a few hundred KB as HEIC. The shot saved to Photos stays full size.

Location is only used while the app is on screen: it geotags a catch and tells HUNT and the camera what's nearby. There's no background location mode at all. Lock Screen tracking sends the proxy the vehicle's number, the stretch of its route up to your stop, and the token it pushes updates to. It's forgotten when tracking ends.

Debug mode (Settings) saves every shot to Files › On My iPhone › TABOR › Debug, including failed reads and retakes. Each shot is saved with the photo, the sticker and what the OCR, the live feed and the sticker cutter made of it. With debug mode on, long-pressing a caught sticker in the book deletes it.

---

Tests

The matching, OCR filtering, stats, badges, live feed parsing, routes and backup logic in Tabor/Core doesn't depend on the iPhone. It also builds as a Swift package, so it's tested on a Mac in about a second:

swift test

The proxy's tracking maths has its own:

node proxy/src/trackgeo.test.mjs

GitHub Actions runs the Swift tests on every push to main and every pull request that touches Tabor/Core, the tests or fleet.json.

---

Thanks

To everyone who tested the beta, and above all to @Szpieg0, @Pepe-PL and @flyingkonstal, whose issues found the bugs, named the vehicles and shaped a lot of what's here. Want to help too? Start with CONTRIBUTING.md.