clamepending/thread-in-a-ropestack

Thread in a Ropestack

Long, multi-camera, static-CCTV video from deterministic simulations, with questions generated from the simulation truth. Like needle-in-a-haystack, but the "thread" is information spread over time, cameras and space: how long something lasted, how often something happened, what one person did across several views, what did not happen.

Code, simulator and renderer: https://github.com/Clamepending/thread-in-a-ropestack-eval

Episodes

One short and one long episode per venue, named <venue>_<minutes>min.

episode	venue	minutes	cameras	fps	resolution	questions	no-video guesser
cornerstore_5min	cornerstore	5	6	15	1280x720	21	0.12
cornerstore_30min	cornerstore	30	6	15	1280x720	62	0.15
nightowl_5min	nightowl	5	7	15	1280x720	22	0.12
nightowl_30min	nightowl	30	7	15	1280x720	59	0.21
eastgate_10min	eastgate	10	8	15	1280x720	32	0.14
eastgate_30min	eastgate	30	8	15	1280x720	64	0.17
riverside_10min	riverside	10	10	15	1920x1080	31	0.14
riverside_30min	riverside	30	10	15	1920x1080	61	0.15
grocery_scan_5min	grocery_scan	5	3	15	960x540	30	0.17

This is release v1.1.1 (382 questions). Earlier releases stay available by tag (revision="v1.0" etc.); see Versions.

Venues

Five venues, deliberately different domains (indoor and outdoor; layout, lighting, textures, camera placement, lens and recorder look). The three indoor venues also have incidents that leave a visible trace for a while: spills until they are mopped up, wet-floor signs, baskets left on the floor, litter bins filling up and being emptied. The outdoor mall has vehicles, a dog and a utility crew instead.

cornerstore

A daytime grocery: front door, gondola aisles, two checkouts, a stockroom. A shopper drops a carton of juice (the stocker puts out a wet-floor sign and mops it up) and another leaves their basket in an aisle and walks out.

Example (intervals). The customer who took an item from the apples shelf at 15:06:29: during which periods were they inside the store?

Answer: 15:06:10.7 to 15:06:50.6

Evidence (from key.json): 15:06:29 on market: takes apples; 15:06:10 on door: walks in; 15:06:50 on door: walks out.

nightowl

A late-night city bodega: side door, coolers, wooden islands, register at the back, watched by seven standard colour cameras. Someone spills a can of soda; someone abandons a basket.

Example (choice). Someone spilled something on the floor. Who caused the spill?

• man in a grey-green shirt and beige chinos
• woman in a purple tank top and black trousers
• man in a light blue shirt (correct)
• bearded man in a white robe and a beige cap

Evidence (from key.json): 23:34:00 on front: drops a can of soda; 23:34:02 on front: spill on the floor.

eastgate

A commuter rail station at rush hour: ticket machines and window, kiosk, fare gates, a platform where trains arrive, open their doors and leave, plus a fisheye dome camera over the concourse. Commuters throw rubbish into three bins that fill up and are emptied by a cleaner, who also mops up spilled drinks. Background people may share outfits; questions only name people whose look is unique in the episode, and otherwise only count them.

Example (choice). Which of these is what the man in a red long-sleeved top and dark jeans did at the station, in order?

• bought something at the kiosk, then bought a ticket at the window, then went through the fare gates (correct)
• bought a ticket from a machine, then bought something at the kiosk, then went through the fare gates
• bought something at the kiosk, then bought a ticket from a machine, then went through the fare gates
• bought a ticket at the window, then bought something at the kiosk, then went through the fare gates

Evidence (from key.json): 07:39:28 on kiosk: bought something at the kiosk; 07:39:53 on concourse: bought a ticket at the window; 07:40:04 on gates_in: went through the fare gates.

riverside

An open-air strip mall and its parking lot: five shops (bakery, pharmacy, hardware, bookshop, cafe), 54 bays, a bike rack and a street with power lines, watched by ten outdoor cameras. People arrive in cars (every car looks different) or on bikes, visit shops and drive out through the exit; someone leaves their dog waiting outside the cafe; a utility crew parks by a power pole, sets up cones and a ladder, and works on it while their truck screens a car parked behind it from the street side. Questions link people to the vehicles they came in.

Example (choice). Which shops did the woman in a green T-shirt and light jeans (who arrived in the dark red minivan) go into, in order?

• Hardware, then Corner Cafe
• Riverside Bakery, then Page & Co. Books
• Riverside Bakery, then Pharmacy
• Pharmacy, then Riverside Bakery (correct)

Evidence (from key.json): 11:02:47 on lot_west: parks; 11:02:50 on lot_west: gets out; 11:03:03 on shops_west: goes into the pharmacy; 11:05:46 on shops_west: goes into the riverside bakery.

grocery_scan

A five-minute pilot in the produce aisle of a scanned real grocery store, watched by three cameras. Two freestanding sample tables (apples, bunches of bananas) and a stack of backstock crates are inserted into the aisle; shoppers take produce and staff refill the tables from the crates. There is no checkout. Questions are about the tables' stock over time and who took what.

Example (choice). In order, which tables did the man in a red long-sleeved top and dark jeans take items from?

• apple table, then banana table (correct)
• banana table
• banana table, then apple table
• apple table

Evidence (from key.json): 15:00:23 on produce_tables: takes apple; 15:00:35 on produce_back: takes bunch of bananas.

Task

A model watches every stream of an episode (store clock burned into each frame) and then answers the questions in task.json, replying with one JSON object mapping question id to answer.

Each videos/<camera>.mp4 holds exactly the rendered frames at the episode's rate (the fps column above): frame $k$ shows episode time $k/\text{fps}$ s, and task.json's fps and frames are read from the files themselves. From v0.8 cornerstore and nightowl are rendered at 15 fps, a common CCTV recording rate; from v1.0 every episode is 15 fps. Since v0.7 no episode is below 4 fps (v0.6.1 to v0.6.3: 2 fps), so a walking person's box overlaps their box in the next frame and overlap-matching trackers keep their tracks. (Up to v0.6 the MP4s repeated each 2 fps image five times at 10 fps while task.json said 2 fps; always read the timing from the container.) Every question has one of five answer types:

answer type	reply with	scored
count	a whole number	by relative error: $\max(0, 1 - 2e)$
duration	seconds	by relative error: $\max(0, 1 - 2e)$
intervals	[[start, end], ...] store-clock times	temporal IoU of the union of periods
moments	[time, ...], [] if it never happened	matched Jaccard: times pair one-to-one within 5 s
choice	one of the options	1 if right

key.json holds the answers plus, for every question, the evidence moments (time, entity, cameras that show it) and answer regions; questions are kept only if every person evidence moment is visible in the rendered pixels.

Scoring

Each answer scores $s_i \in [0, 1]$; unanswered or unparseable answers score 0. Times are the clock burned into the footage, converted to seconds. With $\tau$ the question's tolerance (5 s):

count, duration (relative error $e = |\hat y - y| / y$): $s = \max(0,\ 1 - 2e)$, so 1 when exact, 0.8 at 10% off and 0 from 50% off; when the true value is 0 the answer must be exactly 0.

choice: $s = \mathbb{1}[\hat y = y]$ (compared case-insensitively).

intervals (temporal IoU): merge overlapping periods into unions $\hat U$ (answer) and $U$ (truth), then

$$s=\frac{\mathrm{\mid }\widehat{U}\cap U\mathrm{\mid }}{\mathrm{\mid }\widehat{U}\cup U\mathrm{\mid }}s\; =\; \backslash frac\{|\backslash hat\; U\; \backslash cap\; U|\}\{|\backslash hat\; U\; \backslash cup\; U|\}$$

measured in seconds.

moments (matched Jaccard): pair answered times $\hat T$ with true times $T$ one-to-one, a pair only if $|\hat t - t| \le \tau$; with $m$ the size of the largest such matching,

$$s=\frac{m}{\mathrm{\mid }\widehat{T}\mathrm{\mid }+\mathrm{\mid }T\mathrm{\mid }-m}s\; =\; \backslash frac\{m\}\{|\backslash hat\; T|\; +\; |T|\; -\; m\}$$

Extra or missed moments both cost, and answering at the wrong times scores 0.

For both intervals and moments, an empty answer when nothing happened scores 1 ($0/0 := 1$).

Episode score: the mean over its $N$ questions, $S = \frac{1}{N}\sum_i s_i$. Per-answer-type and per-skill means are reported alongside. scores.json gives two reference points: the answer key itself (oracle, always 1.0) and a guesser that never sees the video (always 3, 60 s, no periods, never, and the first option).

Skills

Independently of its answer type, every question is tagged in key.json with the skills it needs, from a fixed vocabulary. Some questions need exactly one skill (single-skill probes, e.g. "who was paying at the checkout at 15:10:20?" needs only identify, since the time and place are given); most combine several (e.g. "how long after going through the gates did the man in a red sweater get on a train?" needs identify, track and localize-in-track). python -m ropestack score reports the mean per skill, and for each skill the mean over its single-skill questions (only) and over composites (with_others): a method that does well on each skill alone but poorly in combination has trouble composing them.

skill	what it takes	questions	single-skill
find-event	find when a described event happens (no particular person to follow)	96	9
identify	pick out a person from a description of their clothes, or describe who did something	137	12
track	follow one person through time and across cameras	179	6
localize-in-track	find a particular moment within one person's activity	120
count	count distinct people or occurrences without double-counting	96	30
state-over-time	tell when, or for how long, something was true	77	16
snapshot	describe the scene at one given moment	27
order	put events in the order they happened	43
absence	notice something that went wrong or did not happen	7
associate	keep hold of one person through a look-alike or a stretch out of view (a description match is not enough)	13
cross-camera	no single camera shows all the evidence (added automatically)	64

Files

episodes/<id>/task.json            model input: streams + questions (no answers)
episodes/<id>/key.json             answers, tolerances, evidence, regions, skills, difficulty
episodes/<id>/scores.json          oracle (must be 1.0) and no-video guesser references
episodes/<id>/questions.json       all generated questions, including ones dropped as not observable
episodes/<id>/videos/<cam>.mp4     camera streams (clock burned in)
episodes/<id>/annotations/<cam>.jsonl   per frame: time and each visible person's box and pixel count
episodes/<id>/masks/<cam>.tar      per-frame person ID masks (PNG)
episodes/<id>/truth/               events, per-person tracks, shelf stock, intervals, cast, meta
episodes/<id>/scenario.yaml        scenario (intent + seed) the episode was simulated from
episodes/<id>/venue.json           floor plan, fixtures, products, cameras

Scoring code: python -m ropestack score episodes/<id> response.json in the repository.

Provenance and licences

Everything is simulated and rendered (Unreal Engine 5.8); no real people or places. Assets are AI-use-compatible: Microsoft Rocketbox avatars and animations (MIT), CMU motion capture, Poly Haven and ambientCG (CC0), and procedurally generated fictional brands, signs and surfaces.

Versions

Tags mark releases (v0.1, v0.2, ...; from v0.3 one short and one long episode per venue; v0.4 adds the incidents and the station's dome camera; v0.8 renders cornerstore and nightowl at 15 fps, adds anchored questions that name a customer only by one act at a stated time, and scores counts and durations by relative error).

• v1.0: 8 episodes, 352 reviewed questions, every episode at 15 fps.
• v1.1: every episode re-rendered with a realism pass. The pass adds motion-captured actions (picking up, carrying, mopping, sitting into and out of cars), two-handed props (mop and bucket, baskets), modelled packaged goods on shelves and in baskets, car doors that open, a kiosk register and card readers, and real-world sizes. Riverside is re-simulated (nobody walks through cars or bikes, purchases leave in bags; 31 + 61 reviewed questions) and nightowl starts with full shelves. Cornerstore, nightowl and eastgate keep v1.0's questions and answers. grocery_scan_5min is added.
• v1.1.1: eastgate re-rendered with ticket machines modelled on real station machines and with a carried snack bag held in the hand; questions and answers unchanged. The other episodes are identical to v1.1.

Each episode's truth/meta.json names the scenario version it was simulated from and carries a sha256 of the truth, so a re-simulation can be checked against the release.