phi-lab-rice/GRADE_Dataset

GRADE dataset

This dataset release contains raw recordings for GRADE: Single-Frame Generative Radar Depth Estimation Under Visual Degradation, accepted to ACM MobiCom 2026. See the paper and project website.

The recordings were made with a TI IWR1843BOOST radar (with DCA1000EVM), a ZED 2i stereo camera, and a DJI Action 5 Pro camera. Evaluation sequences also used a MAX30105 smoke sensor. All 70 public sequences now include anonymized ZED and DJI RGB videos. The original ZED SVO2 and DJI recordings are not published.

Folders and splits

GRADE_Train_Raw/       50 sequence folders
GRADE_Eval_Raw/        20 smoke evaluation sequence folders
processing_code/       processor.py, requirements.txt, and utils/

Split rule: Within GRADE_Train_Raw, the 10 folders with Dell or Keck in their names are for validation, not model training. This includes the Smoke-Dell-* and Smoke-keck-* folders. The other 40 folders are for training. All 20 folders in GRADE_Eval_Raw are for evaluation. Do not treat every folder under GRADE_Train_Raw as training data.

Each sequence folder contains these timestamp and radar files:

radar_*.h5                 raw radar samples and timestamps
camera_timestamps_*.h5     ZED timestamps
dji_timestamps_*.h5        DJI timestamps
max30105_*.h5              MAX30105 samples and timestamps (evaluation only)

Training sequences have no MAX30105 file. The processor handles that automatically.

Each sequence folder on the Hub also has:

zed_video_anonymized.mkv   calibrated rectified-left ZED RGB, defaced, 1280x720
dji_video_anonymized.mkv   fisheye-corrected and cropped DJI RGB, defaced, 1280x720

Both videos use lossy H.265/HEVC Main profile at CRF 14, with 8-bit YUV 4:2:0 pixels in Matroska. Faces were blurred before this encoding. The original anonymized lossless RGB files were used as the transcoding inputs. Three original DJI recordings (brk-2, brk-3, and brk-basement) have 8, 3, and 6 untimed terminal video frames, respectively. Their anonymized videos omit those frames during CRF encoding, so every released DJI frame has a matching entry in its original DJI timestamp H5. No timestamps were removed from the H5 files.

The SDK-free processor also requires zed_depth.h5: full-frame ZED depth aligned with the ZED RGB video at 1280×720. This depth file has not yet been uploaded to the Hub. The local prepared file has a depth_mm dataset of shape (frames, 720, 1280), dtype uint16, with no HDF5 compression, matching the radar H5 compression setting. Zero marks invalid depth; positive values are millimeters rounded to 1 mm. Depth was computed from the original SVO2 before RGB face blurring. The ZED camera clock is camera_timestamps_*.h5; there is no separate SVO2 timestamp file. Frame index i in that H5 corresponds to frame i in the ZED video and depth H5, up to the exported frame count. The processor trims surplus terminal timestamps to that count.

The local preparation folders retain the original DJI recordings and ZED SVO2 files for preservation. Do not include those originals in a public upload.

Process a sequence

processing_code/processor.py first matches the H5 sensor timestamps, then extracts synchronized radar, ZED RGB/depth, and DJI RGB frames. It saves one output folder per sequence. processing_code/utils/ contains the sensor extractors, radar processing, synchronization, and point-cloud code. The processor requires both anonymized videos and the matching zed_depth.h5; because the depth H5 is not on the Hub yet, the current public download alone cannot run the full camera pipeline.

1. From this repository's root folder, create a Python environment and install the dependencies:

   python3.12 -m venv .venv
   source .venv/bin/activate
   python -m pip install -r processing_code/requirements.txt
   

   The public processor does not require the ZED SDK or an NVIDIA GPU. It uses OpenCV to decode the HEVC videos and h5py to read precomputed metric depth.

2. Process one evaluation sequence after its prepared camera files are available:

   python processing_code/processor.py \
     --dataset GRADE_Eval_Raw \
     --sequences Smoke-keck-2-3 \
     --output-dir processed_eval
   

For a training sequence, use --dataset GRADE_Train_Raw --sequences AlicePratt-1. Leave out --sequences to process every folder in the chosen dataset.

The default output for one evaluation sequence looks like this:

processed_eval/Smoke-keck-2-3/
├── radar.npy            complex64 radar spectrum, (N, 64, 2, 8, 256)
├── dji_rgb.npy          uint8 RGB, (N, 504, 896, 3)
├── zed_rgb.npy          uint8 RGB, (N, 504, 896, 3)
├── zed_depth.npy        uint16 depth in millimeters, (N, 720, 1280)
├── max30105.npy         uint32, 3 channels; only when the sensor file exists
├── sync_triples.csv     original radar, ZED, and DJI frame indices
├── metadata.json        shapes, timestamps, and processing settings
└── pcd/
    ├── pcd_0.npy        float32 XYZ points, (number_of_points, 3)
    └── ...

Each point cloud is a separate .npy file because its number of points varies by frame. N is the number of synchronized frames kept for that sequence.

Useful options

Option	What it does
--dataset PATH	Choose GRADE_Train_Raw or GRADE_Eval_Raw (required).
--sequences NAME [NAME ...]	Process selected sequences; omit to process all.
--output-dir PATH	Choose where outputs go (default: processed).
--no-radar, --no-camera, --no-dji	Skip the named processed output. --no-radar skips the spectrum; use --no-pcd to skip point clouds. All three sensor timestamps are still used for synchronization.
--no-pcd	Skip point clouds (they are produced by default).
--no-doppler	Save radar_no_doppler.npy instead of the full radar.npy.
--list-unprocessed	List sequences without an output metadata.json.

--rgb-codec is a legacy argument and has no effect on the current .npy output. Run python processing_code/processor.py --help for the complete CLI help.

Please cite GRADE if you find this dataset useful in your research

@inproceedings{zhao2026grade,
  author    = {Bin Zhao and Patrick Chiou and Nakul Garg},
  title     = {{GRADE}: Single-Frame Generative Radar Depth Estimation Under Visual Degradation},
  booktitle = {Proceedings of the 32nd Annual International Conference on Mobile Computing and Networking},
  series    = {MobiCom '26},
  year      = {2026},
  publisher = {ACM},
  address   = {New York, NY, USA},
  numpages  = {15},
  doi       = {10.1145/3795866.3844478},
  url       = {https://doi.org/10.1145/3795866.3844478}
}