Datasets:

kaichen-z
/

RISE

pretty_name: RISE-data-4scenes
tags:
  - radar
  - mmwave
  - indoor-scene-understanding
  - layout-reconstruction
  - object-detection
task_categories:
  - object-detection
size_categories:
  - 1K<n<10K

RISE-data-4scenes

Radar frames for four real indoor scenes of the RISE benchmark (RISE: Single Static Radar-based Indoor Scene Understanding, CVPR 2026, arXiv 2511.14019), packed in the python-native bame profile that the released code reads directly. No MATLAB is needed.

Code: https://github.com/kaichen-z/RISE · Project page: https://rise-cvpr.github.io/

case	scene	trajectories	background	frames (traj / bkg shipped)
19	`Raw_Data_3/scene3_set2`	`_r1 _r2 _r3`	`scene3_set2_bkg`	1,108 / 13
26	`Raw_Data_5/scene5_set1`	`_r1 _r2 _r3`	`scene5_set1_bkg`	1,003 / 13
42	`Raw_Data_9/scene9_set1`	`_r1 _r2 _r3`	`scene9_set1_bkg`	1,127 / 13
50	`Raw_Data_11/scene11_set1`	`_r1 _r2 _r3`	`scene11_set1_bkg`	1,103 / 13

Every scene has furniture; the evaluation ground truth for all four cases ships with the code (5_diffusion/data/GT/{obs,wall}/<case>_gt.png).

Download

hf download kaichen-z/RISE RISE-data-4scenes.zip --repo-type dataset --local-dir .
unzip RISE-data-4scenes.zip
sha256sum -c RISE-data-4scenes.zip.sha256     # optional

Layout

RISE-data-4scenes/
  Raw_Data_3/
    scene3_set2_r1/preprocess/
      original_mat_1_<n>.npy      complex64, shape (256 range, 1 doppler, 16 RX, 12 TX), 393,216 B
      original_hyp_1_<n>.json     range-FFT parameters for that frame (7.3 KB)
      python_frames.json          {"profile": "bame", "every": 1, ...}
    scene3_set2_r2/  scene3_set2_r3/            same
    scene3_set2_bkg/preprocess/                 empty-room reference, every 30th frame only ("every": 30)
    *.ply                                       lidar scans of the room (reference only, not read by the code)
    Cascade_Capture_22xx.{mmwave,setup}.json    radar configuration (provenance only)
  Raw_Data_5/  Raw_Data_9/  Raw_Data_11/        same structure

Point the code at the unzipped folder and run a case:

export RISE_DATA_ROOT=/path/to/RISE-data-4scenes
export RISE_DEVICE=cuda            # cuda | mps | cpu (3D CFAR needs a GPU)
cd 3_bame      && python run.py --case 42
cd ../4_inversion && python run.py --case 42 --step associate && python run.py --case 42 --step layout

Format notes

bame profile = doppler bin 0 of the post-MATLAB cube, stored as complex64. This is exact for the released BAME pipeline (3_bame/bame.py slices matrix[:, :1] before a per-column range FFT); it is not sufficient for the MVDR baseline, which needs all 16 doppler bins.
Frame index <n> is the capture frame number; trajectory captures start at 0, background captures at whatever frame the recording started on. The code sorts by <n>.
Packed with tools/pack_bame.pl (stdlib Perl), verified byte-identical to tools/pack_frames.py --profile bame on the same frame.
Raw .bin captures and the MATLAB .mat cache are not included.

Citation

@inproceedings{zhou2026rise,
  title={Rise: single static radar-based indoor scene understanding},
  author={Zhou, Kaichen and Dodds, Laura and Afzal, Sayed Saad and Adib, Fadel},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={32194--32205},
  year={2026}
}