Indoor Radio Map Dataset

Dataset for the ICASSP 2025 First Indoor Path Loss Prediction Challenge, more info can be found at: https://indoorradiomapchallenge.github.io/

The zip file includes (also see the documentation; note that the test_data include only the inputs):

(i)  An "Input" folder containing an indicative input tensor for each simulation (normal incidence reflectance & transmittance for the geometry & distance)

(ii) An "Output" folder with the PL radio map for each simulation 

(iii) A "Position" folder including the Tx positions for all buildings, frequencies, and antennas

(iv) A "Radiation_Patterns" folder comprising the azimuth radiation pattern (for simplicity simulations assume no elevation) for each one of the 5 antennas (Ant1 is isotropic, so all gains are 0)

(v) A "Building_Details" folder including the geometry boundary (most useful info here is W, H, but one can also get these from the image size)

Naming Conventions: B(#ID)_Ant(#ID)_f(#ID)_S(#ID)  ~ for buildings (#ID) ranges from 0 to 25, for the antennas from 1 to 5, for frequencies from 1 to 3, and for samples from 0 to 49 (for Ant1) and 0 to 79 for (Ant2, Ant3, Ant4, and Ant5)

The script file includes indicative code snippets for:

(i) Reading a sample and getting the sample frequency. Using additional features related to frequency will be helpful for Tasks 2 & 3

(ii) Computing angles and extracting antenna gain; you can use this information to extract additional features for Task 3

(iii) Resizing data to a standard scale (note that downsampling substantially compared to the initial dimensions might distort the geometry)

(iv) Augmenting data by flipping. We advise further exploring data augmentation techniques for your work, e.g., additional rotations, cropping, etc.

(v) An indicative data loader in torch

Notes: 

1. If you use data augmentation techniques you can modify your loader accordingly or save the augmented data as new images

2. If you extract new features you can modify your loader accordingly (to perform additional operations) or save the new input tensors as new images. For instance, a 4-channel tensor can be saved as an RGBA image  with the same naming convention, or if you have more channels you can create images with the same name convention with two parts, P1 (channel 1 to 4) and P2 (channel 5 to 8), that you concatenate when you read the data.