# Kitchen CCTV Monitor Challenge - Round 1

Status: live
Prize: $300  
Round: July 12-September 9, 2026
Public page: https://builderr.ai/kitchen-video

## Task

Build an agent that answers structured operations questions from fixed-camera kitchen or restaurant videos.

Round 1 uses public sample videos and public sample questions for orientation. Scoring uses hidden questions and an answer key on those videos. If held-back rights-cleared clips are added later, that will be stated before they are used.

The real use case is simple: a restaurant or cloud kitchen already has CCTV, but no one has time to review an hour of footage. The agent should help answer practical questions: did staff wear caps or hairnets, did food sit too long at handoff, was a container sealed, where did orders back up, and what was not visible from the camera.

Input:

- Video files or video URLs supplied by the evaluator.
- A JSON file of questions for each video.

Output:

- One JSON file with answers.
- One log file showing runtime, frame sampling, model/API calls, and estimated cost.

Example run:

```bash
python answer.py --videos ./videos --questions questions.json --out answers.json --log run_log.json
```

## Interface

`questions.json`:

```json
[
  {
    "id": "q001",
    "video_id": "sample_01",
    "type": "yes_no",
    "question": "Did the worker close the container before moving it away from the station?"
  }
]
```

`answers.json`:

```json
[
  {
    "id": "q001",
    "answer": "yes",
    "confidence": 0.72,
    "evidence": [
      { "video_id": "sample_01", "timestamp_start": 91.2, "timestamp_end": 94.8 }
    ]
  }
]
```

`run_log.json`:

```json
{
  "runtime_seconds": 812,
  "frames_processed": 642,
  "model_calls": 38,
  "estimated_model_api_cost_usd": 0.18,
  "normalized_model_api_cost_per_60min_usd": 0.18
}
```

## What The Questions Test

The task is not freeform video captioning. It is operational QA:

- Did a required step happen?
- Who did what, and when?
- How many visible objects or people were present?
- Was an item sealed, moved, spilled, left unattended, or handed off?
- Was a cap, hairnet, glove, or sealed container visible?
- Was the question impossible to answer from the footage?

Answer types:

- yes/no
- multiple choice
- count
- timestamp
- duration
- short structured object
- not_visible

## Budget

Hard limits for the full scored run:

- $0.30 estimated model/API cost per 60 minutes of source video.
- 25 minutes wall-clock.
- About 1,500 sampled frames per 60 minutes, or an equivalent frame budget stated by the evaluator.
- One reproducible command.
- No manual inspection during evaluation.

If the scored set is longer or shorter than 60 minutes, the dollar and frame caps scale with video minutes unless the evaluator states a fixed cap.

Open-source or local models are recommended because this should be cheap enough for a small kitchen to run often. They are not mandatory. Cloud models are allowed if every call is logged and counted. Local runs count as $0 model/API cost, but still have to fit the runtime and frame limits.

Submissions that exceed the hard budget are ineligible. Lower normalized cost only helps among valid runs with useful accuracy.

## Scoring

Total: 100 points.

- 80 points - answer accuracy on hidden questions.
- 15 points - cost efficiency among valid runs.
- 5 points - reproducibility and auditability.

Tie-breaks:

1. Higher answer accuracy.
2. Lower normalized model/API cost per 60 minutes.
3. Faster wall-clock runtime.

Timestamp answers:

- Full credit within 2 seconds.
- Partial credit within 5 seconds.
- Zero beyond that, unless the event span is long and the rubric says otherwise.

## Sample Questions

These examples show the shape. Scoring uses hidden questions on the same public videos.

- At what timestamp was the first sealed bag placed on the handoff shelf?
- Was the cook at the stove wearing a cap or hairnet at 00:45?
- How many people were active at the prep counter at 00:45?
- Did the worker close the container before moving it away from the station?
- Which happened last: garnish added, lid closed, bag moved, or tray wiped?
- Is the order number visible? Answer `not_visible` if it is not readable.

## Benchmark To Beat

Before ranking final submissions, builderr should run a simple baseline:

1. sample the whole clip cheaply,
2. identify likely time windows for each question,
3. run a common vision-language model or video model only on those windows,
4. run OCR where relevant,
5. answer each question in JSON with evidence spans,
6. write a frame/model-call/cost log.

The benchmark is meant to be real but beatable. Broad captions or one-pass summaries should lose on timestamps, counts, event order, and `not_visible` cases. The winning submission should beat the baseline while staying under the $0.30 per-hour cap.

## Current Public Video Sources

These are public sample and reference leads, not a hidden test set:

- 60-minute fixed kitchen CCTV-style service clip: https://www.youtube.com/watch?v=eAXIkGQYYWQ
  - One-hour fixed kitchen view with timestamp, staff movement, prep, and service flow. Strong public sample/reference; verify rights before rehosting.
- Cctv dapur keteter pesanan mulai jam 2 sampai jam 8 pagi: https://www.youtube.com/watch?v=ZrlbDd4Bs2Q
  - About 28 minutes of busy kitchen order flow from a fixed CCTV-style angle. Useful for hidden-question design and cost-control testing.
- CCTV DAPUR: https://www.youtube.com/watch?v=2Bj50mmQeFM
  - About 22 minutes of fixed kitchen cooking footage. Good for long-window sampling and timestamp questions.
- cctv dapur seafood jos gandos: https://www.youtube.com/watch?v=hwz0t3kckpo
  - About 19 minutes of restaurant-kitchen work from a fixed view. Useful as a public long-form reference.
- Chinese Commercial Kitchen Manipulation Dataset Preview: https://huggingface.co/datasets/nova-dynamics/Chinese_Commercial_Kitchen_Manipulation_Dataset_Preview
  - Fixed-view real commercial-kitchen work. Good supplemental sample source for questions and baseline work.
- Kaggle - Kitchen Video in Restaurants: https://www.kaggle.com/datasets/naoamscoltd/kitchen-video-in-restaurants
  - Restaurant-kitchen workflow footage. License/access must be checked before rehosting cuts.
- Chinese restaurant hygiene-problem footage: https://www.youtube.com/watch?v=mWOoAf4rIhk
  - Messier public restaurant footage. Useful for realism and question design.
- IP Bullet CCTV Camera (Kitchen View) - Revlight Security: https://www.youtube.com/watch?v=nQht56i2Xkg
  - Short true fixed kitchen CCTV. Good visual target, too short to carry the round alone.
- COM Kitchens reference dataset: https://www.nii.ac.jp/dsc/idr/en/rdata/COM_Kitchens/
  - Unedited fixed-view cooking videos; restricted academic access.
- EPFL Smart Kitchen reference dataset: https://github.com/amathislab/EPFL-Smart-Kitchen
  - Long multi-view kitchen actions; useful for benchmark design.

## Notes For Builders

- Do coarse-to-fine search: cheap pass over the whole clip, then inspect likely windows.
- Keep timestamps attached to evidence.
- Use OCR narrowly for labels, receipts, screens, or timers.
- Prefer `not_visible` over guessing when the evidence is missing.
- Log cost and frames. The budget is part of the task.