code_snapshots/radiomap_compare_code/README.md

Radiomap Compare Baselines

This directory stores external comparison baselines for the radiomapvggt project.

Paths:

• code root: /data/jing_code/wl/radiomap_compare
• env root: /data/conda_envs
• dataset roots used for adaptation:
  • /data/jing_dataset/processed/wireless/scannetpp
  • /data/jing_dataset/processed/3D/scannetpp

Paper fixed compare protocol

A fixed paper-facing comparison slice now lives under:

• paper_fixed5_compare

It defines:

• fixed 5 scenes
• fixed centered train/test TX
• fixed centered z order
• reusable scripts for:
  • A06 fixed-case evaluation
  • cross-method metric tables

Current status

1. NeRF2

• repo: NeRF2
• upstream: https://github.com/XPengZhao/NeRF2.git
• local env: /data/conda_envs/nerf2_compare
• env status:
  • created
  • imports verified: torch, yaml, imageio, pandas, skimage, einops
• local notes:
  • readme_compare.md
• current ScanNet++ adaptation status:
  • one real scene export validated:
    • e0d1ed2159
  • exported BLE-format dataset path:
    • /data/jing_code/wl/radiomap_compare/NeRF2/data/scannetpp_ble/e0d1ed2159
  • dataset load verified through upstream BLE_dataset
  • compare scripts/configs added under:
    • /data/jing_code/wl/radiomap_compare/NeRF2/configs
    • /data/jing_code/wl/radiomap_compare/NeRF2/scripts
  • current no-Slurm compare setup also supports:
    • 3 aligned scenes on GPU 5-7
    • automatic train/test postprocess plots

2. Second external baseline repo

2. RF-3DGS

• repo: RF-3DGS
• upstream: https://github.com/SunLab-UGA/RF-3DGS.git
• local env: /data/conda_envs/rf3dgs_compare
• env status:
  • created as a dedicated clone of our known-good gaussian_splatting env
  • verified:
    • torch
    • torchvision
    • diff_gaussian_rasterization
    • repo entry train.py -h
• local notes:
  • readme_compare.md
• current ScanNet++ adaptation status:
  • direct wireless-grid supervision is not plug-and-play
  • repo expects:
    • images/
    • sparse/0/cameras.txt
    • sparse/0/images.txt
    • sparse/0/points3D.ply
    • train_index.txt
    • test_index.txt
  • local exporters now cover both:
    • visual RGB path:
    • /data/jing_code/wl/radiomap_compare/RF-3DGS/dataset_tools/export_scannetpp_visual_to_rf3dgs.py
    • radiomap pseudo-image path:
    • /data/jing_code/wl/radiomap_compare/RF-3DGS/dataset_tools/export_scannetpp_radiomap_to_rf3dgs.py
  • current RF compare now has two modes:
    • visual-only scene fitting
    • radiomap pseudo-image fitting
  • current no-Slurm compare setup also supports a fixed aligned 5-scene run on GPU 0-4, with automatic train/test rendering and overview plotting
  • training-length policy has been updated to match original 3DGS visual training:
    • default iterations = 30000

Comparison strategy

These external baselines are not scene-generalization models like radiomapvggt. They should be used as single-scene fitting baselines.

Recommended protocol:

• choose one scene from the canonical small100 + val10 split
• train one baseline only on that scene
• use the same canonical wireless subset:
  • 27 TX from the centered 3x3x3 subset
  • 3 centered RX z-planes: [4, 5, 6]
• use the remaining scene-local unused data as held-out evaluation when the baseline structure allows it
• compare against radiomapvggt on the same scene

Canonical split reference

Scene split and subset protocol are recorded at the dataset roots:

• wireless README
• 3D README

Notes

• NeRF2 already includes an internal non-neural baseline:
  • mri.py
  • baseline readme_compare.md
• so even before a second external repo is added, we already have:
  • one neural baseline: NeRF2
  • one simple analytic baseline: MRI
• RF-3DGS is structurally different from both of the above:
  • stage 1 is standard 3DGS visual reconstruction from camera images
  • stage 2 expects RF supervision in an image-like / spectrum-like camera frame
  • our current processed wireless dataset is a TX/RX scalar-grid dataset, so RF-3DGS needs an explicit modality conversion layer instead of a direct loader swap
  • the new local radiomap path is exactly that bridge:
    • it exports radiomap slices as pseudo-images
    • it keeps the original 3DGS trainer unchanged
    • it should be interpreted as a forced single-scene fitting baseline