Shoot & source ~70 short test video clips for a computer-vision / security-camera AI system

OVERVIEW We're testing a video-analytics system (object/person detection, safety and security alerts) and need a battery of short evidence clips to validate it. Each feature needs two clips: a positive case where the event clearly happens, and a negative "trap" case that looks similar but should NOT trigger. The negative cases are just as important as the positives. Quality of labeling and adherence to the spec matters more than cinematics. --- SOURCING POLICY (read before anything else) We already hold the standard public CV datasets (RWF-2000, the common face / LPR / PPE / action sets, etc.). Footage from them is worthless to us, we may have effectively tuned on it, so it can't validate anything. Every asset must be net-new. Do not pull from well-known public datasets. Three acceptable routes, your choice per clip: 1. Self-recorded / staged — you film it. Preferred for anything involving real people, motion, or physical events. 2. AI-generated — synthetic video/stills. Good for appearance and static attributes (see limits below). 3. Sourced from the open web or other means — only if you have, and can pass to us, commercial / internal-use rights to the result. Scraping a random clip does not grant you those rights. AI-generation limits Fine for: appearance/attribute and presence-absence tests — clothing type/colour, age/gender sets, PPE on/off, cones, vehicle colour/make stills, "person present vs zone empty." Not acceptable for: anything where physical motion or fine real-world detail is what's under test — falls, fights, gait, vehicle speed, tracking/occlusion, and real LPR. Current video models produce temporal artifacts and implausible motion that can falsely pass a detector or trip it in ways real footage won't. Record these for real. State the route and (for sourced/AI) the tool or source + license for every asset in notes.csv. --- TECHNICAL SPEC (unless a row says otherwise) - Format: MP4 (H.264) for video; JPG/PNG for stills - Resolution: 1920x1080 minimum - Frame rate: 25 fps - Duration: 10–30 s per clip (a few longer — noted) - Audio: not required - One subject/event clearly visible and unobstructed Camera rig (so clips and zone definitions are comparable) Ideally you would mount as a fixed surveillance camera would be. Indicatively, unless a row says otherwise: - Height: 2.5–3.5 m - Tilt: 15–30° down from horizontal - FOV: wide, roughly 90–110° horizontal (approx. 2.8–4 mm equivalent) - Subject distance: primary action 3–8 m from camera - Static — no handheld except DET-TAMP. Note the actual height/tilt/distance per clip in notes.csv. Keep these constant within a POS/NEG pair. For rule-based clips (zones/lines/paths): you can tell us the zone pixel coordinates and frame the action into it. Conditions Default is clean: daylight, unobstructed, single subject. Decide with us up front whether adverse conditions (low light, IR/night, backlight/glare, rain, partial occlusion) are in scope — these are where detectors actually fail, and they matter most for LPR and face, which are largely night-time in deployment. If out of scope for this round, flag it; don't silently ship daylight-only. --- FILE NAMING (strict — this is how we match clips to tests) - Filename = exact Asset ID + extension, e.g. DET-FIRE-01.mp4, DET-CLOT-01a.jpg - -01 = positive (event present). -02 = negative (event absent). This holds for every row, including the PPE rows below — ignore any POS/NEG ordering in the prose, the number is authoritative. - Image-set members: append a lowercase letter, e.g. DET-CLOT-01a/01b/01c, DET-GEND-01a...01d. - Deliver in group folders or flat. Include notes.csv (columns: id, route, source_or_tool, license, camera_height_m, tilt_deg, caveats). --- THE SHOT LIST POS = event clearly present (-01). NEG = similar-looking scene where it must be absent (-02). Identification & vehicles DET-FACE-01/02 — POS: person walks in, looks at camera ≥3s. NEG: different person, similar build/clothing. Provide 3–5 reference stills of the POS person (DET-FACE-gallery/). Consented self-record or synthetic only. DET-LPR-01/02 — POS: vehicle toward camera, plate readable ≥2s, daylight, plate ≥80px wide. NEG: plate blurred / wrong angle / out of view. Consented or synthetic plates only; record real if plate legibility is under test. DET-VCOL-01/02 — stills: red car on neutral background; silver/grey car under yellow sodium streetlight. DET-CONT-01 — still of a shipping container, ISO code visible. DET-VMAKE-01 — clip of a vehicle with visible badge/grille. Security events DET-FIRE-01/02 — POS: visible open flame ≥5% of frame (campfire / controlled burn — a candle at surveillance distance won't reach 5%; if you can only safely do a candle, tell us and we'll relax the threshold for that clip). NEG: red/orange that is NOT fire (sunset, neon, brake lights, cigarette ember). DET-FALL-01/02 — POS: person walking then falling and lying still ≥3s. NEG: deliberately sitting / lying / yoga / pushup / tying a shoe. DET-VIOL-01/02 — POS: two people fighting (shoving/punching). NEG: hugging, dancing, sports pile-on, high-fives. Record safely or stage with willing participants; no real violence. DET-GUN-01/02 — POS: person holding a replica gun. NEG: person holding a drill / phone / banana. DET-GBR-01/02 — POS: person raising and pointing the replica. NEG: holstered replica, not brandished; or pointing a phone/finger. DET-TAMP-01 — camera covered / spray-painted / rotated by hand (handheld OK here). Safety & compliance (industrial / worksite) DET-GLOV-01/02 — POS (-01): worker bare-handed handling tools. NEG (-02): same scene, gloves on. DET-HELM-01/02 — POS (-01): worker without hard hat. NEG (-02): same worker with it. DET-UNIF-01/02 — POS (-01): same area, street clothes. NEG (-02): worker in required vest/uniform. DET-PSAF-01/02 — POS: person within ~1m of a moving forklift/truck. NEG: person ≥3m clear. Record real — motion/proximity under test. Person & object attributes DET-HAND-01 — hands visible doing a task (typing/lifting). DET-CLOT-01a/b/c — three stills: red shirt / blue jeans / yellow cap. DET-CLOT-02 — still: person in odd/monochrome lighting. DET-GEND-01a–d — 4 stills (2 male, 2 female, varied lighting). DET-AGE-01a–d — 4 stills (child, teen, adult, senior). Synthetic acceptable; if real people, consent required. DET-CARRY-01/02 — POS: person with backpack/shopping bag. NEG: same person, no bag. DET-CUST-01/02 — POS: yellow safety cone in frame ≥3s. NEG: no cone, other industrial gear present. Zone / line / behavior rules (we supply zone/line/path definitions — you provide the action) DET-PROX-01/02 — two people within ~1m ≥5s / ≥3m apart. DET-TRES-01/02 — person enters marked area / walks around but not into it. DET-ZONE-01/02 — object enters zone and stays / enters briefly and leaves. DET-APP-01 — empty zone, then person walks in. DET-DIS-01 — object in zone, then leaves frame. DET-TRAF-01/02 (25s) — vehicle A to B to C / A to B then turns away. DET-OCC-01/02 (30s) — more than 5 people in zone for more than 10s / less than 4 people. DET-LOIT-01/02 (60s) — person stays more than 30s / walks through less than 10s. DET-LINE-01/02 — person crosses a line A to B / opposite B to A. DET-MISS-01/02 (30s) — marked object present then removed / present throughout. DET-ABS-01/02 (60s) — zone empty ≥30s / always has ≥1 person. DET-DTAIL-01/02 — door opens, two people pass on one entry / just one. DET-TAIL-01/02 — two vehicles through a gate less than 1s apart / more than3s spaced. DET-DOOR-01 (60s) — door propped open more than 30s. DET-SUSP-01/02 — specific person appears / different person (needs reference stills like DET-FACE). Consented or synthetic only. DET-RECUR-01 — same person entering a building in 3 separate short clips. Consented or synthetic only. DET-SPEED-01 — vehicle through frame at speed. Log the actual speed and camera-to-path distance (ground truth — see below). DET-MOT-01/02 — static empty scene then person walks in / empty scene with only mild lighting flicker. Tracking UI-TRACK-01 (20s) — one person crosses frame, briefly hidden behind an object, re-emerges. --- LONG ANALYTICS CLIPS — separate engagement These are a different job (sustained, patterned real foot traffic — can't be staged convincingly in 20s pieces) and require ground-truth logs or they validate nothing. Discuss separately; existing controlled-venue or overhead footage you own/license is fine here. DASH-HEAT-01 (~30 min) — sustained foot traffic, some areas busier. DASH-QDWELL-01 / DASH-QCNT-01 (~30 min) — people forming/moving through a queue. DASH-OCC-01 (~60 min) — occupancy rises and falls over time. Ground-truth requirement (analytics + DET-SPEED) For these clips, deliver a timestamped truth log alongside the footage: - Occupancy: per-minute count in the zone. - Queue: each person's entry and exit timestamp. - Speed: vehicle speed (km/h) and camera-to-path distance (m). Without this we're eyeballing output against nothing. --- QA / ACCEPTANCE A clip is accepted when: correct Asset ID; spec met (resolution, fps, duration, angle); the labeled event is unambiguously present (POS) or absent (NEG); POS/NEG pair shot under matched conditions; sourcing route + license recorded in notes.csv; and (faces/plates) consent or synthetic confirmed. Rejected clips (wrong label, ambiguous event, mismatched pair, missing rights, off-spec angle) are redelivered at no charge within 5 days. Payment is per accepted clip / package on acceptance. --- SAFETY & LEGAL Fire, fights, weapons: nothing dangerous or illegal. Controlled/safe sources only (campfire/controlled burn; staged shoving with willing participants; replica/airsoft handled responsibly). When in doubt, don't stage it — tell us. Everything delivered must be yours to license to us — your own footage, properly-licensed/royalty-free, or generated. State source and license for anything you didn't shoot. --- HOW TO APPLY Tell us (1) which groups you can cover, (2) which you'd film / generate / source per group, (3) your rate (per package). Bonus for prior CCTV-style or dataset-style footage.