Discord

Checklist IOAI 2025 At-Home Round · Task 1

Radar

Build a model that marks, for every cell of a radar range-azimuth map, whether a human is present, using six radar heatmaps.

  • Semantic segmentation (binary)

The task

Radar locates objects from reflected signals, but noise and clutter such as trees make detection unreliable. This practice task asks for an AI model that finds where humans are in a radar's field of view.

Each sample is a 7 × 50 × 181 tensor: six normalised heatmaps (static and dynamic range-azimuth, range-elevation and range-velocity maps) and a range-azimuth label map in which 1 means a human is present and 0 means absent.

Given the six heatmaps, the model must output the 50 × 181 label map. Correct human cells count far more than correct background cells. The on-site Individual Contest Radar task later extended this problem to five classes.

Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.

At a glance

You get
Samples stored as .mat.pt tensors of shape 7×50×181 (six heatmaps plus a binary label map); training and validation sets downloadable from Google Drive as Millimeter-wave_dataset.zip.
You submit
A 50×181 binary map per sample. The baseline notebook submits submission.zip containing the model definition (submission_model.py) and its weights (submission_dic.pth).
Scoring
Weighted accuracy: each correctly identified background point earns 1 point and each correctly identified target point earns 1,500 points; the total is normalised to 0-1 by the maximum possible score. (The repo's scoring script gives 5,000 points per target point instead.)
Format
At-home round (non-graded practice, 1-30 July 2025)

Details

Year
2025, Beijing, China
Round
At-Home Round · Task 1
Language
English
License
CC BY 4.0, as stated by the source