Try to confirm the data preprocessing process

[NeverlandCookies]

Hello @Toytiny and @Ericwen2001

I am studying this repo. And I encountered some problems regarding the data preprocessing. As mentioned in the README, we should run
python generate_4d_polar_doppler.py to apply Data volume reduction( Doppler bins descriptor and Sidelobe-aware spatial sparsifying). As I went on studying the code, I got a few questions.

1. In generate_4d_polar_doppler.py, data = np.load(file_path) # We have already convert .mat to .npy mentioned tesseract_xxx.mat has been converted to npy file. I would like to know whether this is a simple transformation from .mat to .npy? Or is there any operation that I didn't recognised?
2. Regarding the top-Nr mean power value for Sidelobe-aware spatial sparsifying, I noticed that there is a similar operation in extract_pts_feat in radarocc_self.py?
3. I would like to know how n_ranges = 175 in extract_pts_feat in radarocc_self.py is determined, as I couldn't find related range bin filtering in ROI. I couldn't understand the operation of power_val = power_values[batch_idx][:,:original_k*n_ranges]

I am really grateful for your kind reply.

[Ericwen2001]

Hi,

1. Yes. We use .npy files just because they load faster than .mat format.

2, 3. The correct n_ranges calculation should be based on the ROI of [0, 51.2m] and radar resolution of 0.46m. The proper formula is sqrt(51.2² + 25.6²)/0.46 ≈ 125 (this covers the diagonal distance across the ROI). I think the 175 was an error in our implementation. While larger n_ranges might help predictions within ROI, we didn't test this experimentally at the time. Thanks for point out this.

After preprocessing with data volume reduction, the shape of power_val becomes [doppler, k * num_range_bins], where:
k = number of top values selected per range bin
num_range_bins = 256 (total range bins in original data)
The values are sorted by range index. Since using all 256 range bins may be computationally inefficient for our Region of Interest (ROI), we can truncate the data to only include relevant ranges using power_val[:, :k*n_ranges].

Hope this can help.