Checklist HAIO 2026 Summer National Final · Task 3
Shredded Documents
English title: Darált Dokumentumok
Reassemble shredded documents from barcode-marked strips, extract suspect records and identify people in CCTV frames with a given recognition network to solve a fictional theft.
The task
In the fictional city of Velmora a generative video model has been stolen; it changed hands between two suspects on 20 May 2026 in Old Velmora, and the hidden client "Dr. Schmertz" was present in disguise. The evidence consists of shredded documents, CCTV frames with person bounding boxes, a pretrained citizen-recognition model and a citizen register.
Stage 1 (30 points): reorder the 1,922 vertical strips (20 px wide) into three documents — briefing (62 strips, 1 page), suspect list (124 strips, 2 pages) and citizen list (1,736 strips, 28 pages), separable by background colour — using the 44-px, 11-bit barcode at the top of each strip, which may be blurred. Stage 2 (10 points): extract the suspects' partially redacted records and link each to a citizen/class ID via a shared unambiguous field. Stage 3 (25 points): crop the people from the CCTV frames (the boxes cover the whole body, but the model works on portraits; RGB, 64×64, correct channel order and normalisation) and run the recognition network.
Stage 4 (25 points): find the two suspects who carried out the handover, using the fact that random citizens never come unusually close to each other in the footage. Stage 5 (10 points): identify the mastermind, who is not on the suspect list but is in the citizen register and appears repeatedly in the relevant CCTV scenes. The goal is a reproducible pipeline rather than manual inspection.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
In English
Some of this task's files were published only in Hungarian. SOTA translated that file into English on 16 September 2026.
Read the task description (one-pager) in English
Shredded Documents
English translation by SOTA – AI Community of the Hungarian original. Licensed CC BY-NC-SA 4.0, like the original. Organisers who would like this translation removed can email [email protected].
Magyar MI Diákolimpia (Hungarian AI Olympiad)
Task description · Summer National Selection (Nyári Országos Válogató)
May 2026
[Figure: see the original one-pager.]
1. Introduction
Velmora seems to be a peaceful city: it is characterised by old archives, crowded markets, festivals and a dense camera network. Now, however, the city has found itself at the centre of a particularly delicate affair. According to a report by a research group, a sealed data storage device has disappeared that contained a high-performance generative video model. The model can produce realistic faces and voices, so in the wrong hands it could also be used for blackmail, for manufacturing fake news or for staged political incidents.
According to intelligence, the theft was commissioned by an unknown client who uses the code name Dr. Schmertz. The model changed hands between two suspects, and a courier was then supposed to take it on to the client. Based on the information available, the handover already took place on 20 May 2026, in the Old Velmora district of the city. According to the reports, the real mastermind of the operation may also have been present at the scene, in disguise, blending into the crowd.
The persons, places and institutions appearing in the task are fictitious.
2. The task
Several kinds of evidence are available to you for the investigation: shredded documents, footage from surveillance cameras, the bounding boxes of the persons detected in the images, a pretrained recognition model, and a citizen registry.
The goal is to find out from these data:
- who is on the list of suspects,
- which two suspects carried out the handover,
- who was the client directing things from the background.
3. What you receive
The attached notebook contains the story introduction, the description of the subtasks, some starter code snippets and tips that can help you with the first steps.
The data include the following files:
- images of the shredded document strips,
- CCTV footage,
- the bounding box files belonging to the CCTV images,
- a pretrained recognition model,
- documents containing citizen and suspect data.
The documents were cut into vertical strips, and the strips were then shuffled. The strips are 20 pixels wide. The barcode-like band at the top of the documents can help you decide which page a given strip belongs to and where it may have been located in the original document. The band is not necessarily perfectly clean, so it is advisable to use some more robust similarity measure rather than exact pixel matching.
The persons visible in the CCTV footage have bounding boxes. Your task is to crop out these image regions, prepare them appropriately and then run the provided recognition model on them. From the resulting predictions you can tell which citizen or suspect is visible in each piece of footage. The given bounding boxes outline the full figure, not the portrait. The model works only on portraits.
4. Subtasks and scoring
The maximum score is 100 points.
- Ordering the document strips – 30 points
Restore the shredded documents. The goal is to produce readable pages again from the shuffled strips. - Extracting the suspects' data – 10 points
Extract the names and identifiers of the suspects from the restored documents. These identifiers will later be needed as the class labels of the image recognition model. - Running the recognition model on the CCTV footage – 25 points
Using the bounding boxes, crop out the persons appearing in the images, then give them as input to the pretrained network. Pay particular attention to the input format: the images must be in RGB format and 64×64 pixels, the channel order must match what the model expects, and the normalisation must also be correct. - Identifying the two suspects involved in the handover – 25 points
Based on the model's predictions, find out which suspects are visible in the CCTV footage. Then examine which two suspects came conspicuously close to each other in the same image. The goal of this subtask is to determine between whom the handover took place. - Finding the person directing things from the background – 10 points
According to the investigation, the real client is not on the list of suspects, but is in the citizen registry and also appears in several relevant pieces of CCTV footage. Identify who is hiding behind the code name Dr. Schmertz.
5. Submission
The submission must contain the code used for the solution and the results obtained. It must be clear from it:
- how you restored the documents,
- which suspects and identifiers you extracted,
- how you prepared the CCTV image regions for the model,
- whom the model recognised in each piece of footage,
- which two suspects carried out the handover,
- and who was the person directing things from the background.
The aim of the task is not to look through the images by hand, but to build a re-runnable investigation pipeline. A good solution links, step by step, the information extracted from the documents, the images and the tabular data.
For every subtask, the notebook specifies the format in which the evaluator expects the solutions.
6. Useful resources
Translated by SOTA. The Hungarian original is the official version and wins wherever the two differ. Original by the Hungarian AI Olympiad (ELTE Faculty of Informatics), licensed CC BY-NC-SA 4.0; this is a translation of the task one-pager, and the official English notebook of the task is linked on this page. This translation is shared under CC BY-NC-SA 4.0, the licence of the original. If you organise this olympiad and would like the translation removed, email [email protected] and we will take it down.
At a glance
- You get
data/shreds(1,922 PNG strips),data/cctv(PNG frames with CSV bounding boxes: left, top, right, bottom), three background images, an example reconstructed briefing page with barcode, andbest_citizen_face_model.pt(108 classes). The shreds, backgrounds, example page and model are in the repository folder adatok/daralt-dokumentumok; the notebook downloads the data from Google Drive.- You submit
- One CSV per stage uploaded to DOCK: Stage 1 — id,
first_name,last_name,date_of_birthfor citizens 7, 12, 21, 35 and 42; Stage 2 — the full records of all 7 suspects; Stage 3 — a CSV with an id column listing citizen IDs; Stages 4 and 5 — CSVs with header id,first_name,last_name,date_of_birthwith at most 2 rows and exactly 1 row respectively. - Scoring
- Stage 1: proportional, 15 fields (5 rows × 3 fields) → 30 points. Stage 2: all-or-nothing 10 points. Stage 3: set overlap with the citizens appearing in scenes that contain a suspect — 100% coverage 25 points, at least 50% coverage 15 points, otherwise 2 points per correct ID; 0 points if the submitted set is more than twice the size of the true set. Stage 4: IDs only — 5 points per correct ID (maximum 10); complete records — both correct 25, one correct 10, none 0; more than 2 rows → 0. Stage 5: ID only — 5 points; complete record — 10 points, only ID correct 5, wrong ID 0; more than 1 row → 0. Total 100.
- Rules
- 10 upload attempts per stage with a 30-second cooldown; the best submission counts.
- The seed cell must not be modified.
- Only the provided Windows lab machines may be used; Python is the official language and only the pre-announced list of Python packages is allowed.
- Only the pre-announced whitelist of websites and tools may be used; code-completion and LLM services (e.g. GitHub Copilot, ChatGPT, Claude) are forbidden, except the single AI model named on the whitelist.
- Only computing resources that are completely free at the time of the contest may be used (no paid tiers such as Colab Pro).
- Own notes, books and printed material may be used during the practical part; audio-visual material and all communication are forbidden.
- Mandatory screen recording with OBS Studio.
- Format
- Summer National Final (Nyári Országos Válogató), 30 May 2026, ELTE Lágymányos Campus, Budapest. Practical part: three hours on the provided machines after a short briefing; three 100-point tasks (ML, CV, NLP), scored automatically on the DOCK platform.