CRISP: A Camera–Radar Dataset for Indoor 3D Perception

CRISP (Camera–Radar Indoor Specular-surface Perception) is a synchronized RGB camera, RGB-D, and mmWave radar dataset for indoor 3D perception, focused on environments with specular and reflective surfaces (glass, polished floors, metal) where camera-based depth is missing or unreliable.

• 59,908 frames from 23 predominantly indoor environments (300 takes)
• Two TI mmWave radars: AWR1843 (azimuth) and AWR6843 (elevation)
• Raw ADC data, range–angle maps, and radar point clouds for both radars
• Plane-based geometric annotations for specular surfaces (coming soon)
• 12,094 annotated samples (5 Hz) with two benchmark splits

Paper: CRISP: A Camera–Radar Dataset for Indoor 3D Perception, WACV 2027. Code: https://github.com/ConvergentSystems/CRISP

Dataset structure

CRISP/
├── README.md
├── metadata/
│   ├── frame_index.csv          # maps each frame ID to scene, take (sequence), and side
│   └── splits/
│       ├── split1_take_disjoint.csv
│       └── split2_scene_disjoint.csv
├── calibration/
│   ├── camera_intrinsics.pkl
│   ├── az_rotations.npy         # azimuth radar (AWR1843) -> camera
│   ├── az_translation.npy
│   ├── elev_rotations.npy       # elevation radar (AWR6843) -> camera
│   └── elev_translation.npy
├── images/part_XXX/frame_XXXXXX.png
├── depth_gt/part_XXX/frame_XXXXXX.png
├── radar_raw_az/part_XXX/frame_XXXXXX.npy
├── radar_raw_elev/part_XXX/frame_XXXXXX.npy
├── radar_range_angle_az/part_XXX/frame_XXXXXX.npy
├── radar_range_angle_elev/part_XXX/frame_XXXXXX.npy
├── radar_pc_az/part_XXX/frame_XXXXXX.npy
├── radar_pc_elev/part_XXX/frame_XXXXXX.npy
└── IMU/sequence_XXXX/

• Shards. part_XXX folders are storage shards, not scenes. Frame frame_N lives in part_{N // 5000:03d} (part_000–part_010 hold 5,000 frames each, part_011 holds 4,908). Use metadata/frame_index.csv to find a frame's scene, take, and side.
• Frame IDs. A frame's files share the same global ID (frame_000000–frame_059907) across all modalities.
• Takes and sides. Each take is a unique capture within a scene and is stored as one sequence_XXXX (300 in total). The sensor platform sweeps left-to-right and right-to-left, giving two sides, L and R; most takes have both, and 64 have only one.

metadata/frame_index.csv

One row per frame (59,908 rows).

Column	Description
frame_id	Global frame ID (matches frame_XXXXXX)
sequence_id	Take (sequence_0000 … sequence_0299); also the IMU/ folder name
camera	Sweep side: L or R
local_frame_index	0-based index of the frame within its take and side
camera_frame_index	0-based index into the camera stream for that side (identical to local_frame_index in this release)
part	Shard folder containing the frame
scene_id	Scene identifier (scene-01 … scene-23)

Modalities

All per-frame files are 1280×720 or indexed to that image.

Folder	Format	Shape / dtype	Content
images	PNG	720×1280×4, uint8	RGB frames (stored as RGBA; alpha is always 255)
depth_gt	16-bit PNG	720×1280, uint16	Depth from the RGB-D camera, registered to the RGB frame, in millimetres; 0 = no measurement
radar_raw_{az,elev}	.npy	(8, 128, 512), complex128	Raw complex ADC data (8 virtual antennas × 128 chirps × 512 ADC samples)
radar_range_angle_{az,elev}	.npy	(64, 512), float64	Range–angle magnitude maps (64 angle bins × 512 range bins): mean over chirps, Hamming-windowed range FFT, 64-point zero-padded angle FFT (fft-shifted)
radar_pc_{az,elev}	.npy	(N, 3), float32; N varies per frame	Radar detections projected into the image: u (px), v (px), radar depth (m)
IMU/sequence_XXXX	.npy / .pkl	see below	Gyroscope streams per take

Each IMU/sequence_XXXX/ folder holds, for each side present (left = L, right = R):

• {left,right}_gyro_data.npy — (M, 3) float64 angular rates in rad/s
• {left,right}_gyro_ts.pkl — list of M datetime objects (naive, UTC), one per sample

Loading example

import numpy as np, pandas as pd
from PIL import Image
from huggingface_hub import hf_hub_download

repo = "ConvergentSystems/CRISP"
get = lambda p: hf_hub_download(repo, p, repo_type="dataset")

index = pd.read_csv(get("metadata/frame_index.csv"))
row = index.iloc[0]
rgb   = np.array(Image.open(get(f"images/{row.part}/{row.frame_id}.png")))[..., :3]
depth = np.array(Image.open(get(f"depth_gt/{row.part}/{row.frame_id}.png")))  # mm
pc_az = np.load(get(f"radar_pc_az/{row.part}/{row.frame_id}.npy"))

Calibration

camera_intrinsics.pkl holds the RGB camera intrinsics as a NumPy record with fields fx, fy, cx, cy, width, height (611.34, 611.38, 636.95, 361.54, 1280, 720). The {az,elev}_rotations.npy (3×3) and {az,elev}_translation.npy (3×1) files map radar points into the camera frame: P_camera_mm = R @ P_radar_mm + t_mm, with t in millimetres.

The extrinsics were estimated from corner-reflector captures at ten positions per radar (azimuth: −30° to +30°; elevation: nominally −10° to +10°, at 1.5 m and 3.05 m). The RMS 3-D calibration error is 113.5 mm (azimuth radar) and 101.6 mm (elevation radar), with median reprojection errors of 16.8 and 15.2 pixels.

Benchmark splits

Split files are in metadata/splits/, one row per annotated frame, with columns frame_id, sequence_id, scene_id, side, split (train / val / test).

Split	File	Type	Train / Val / Test
Split-2 (primary)	split2_scene_disjoint.csv	Scene-disjoint: each scene in one partition; test scenes 12, 14, 21, 22	9,596 / 628 / 1,870
Split-1	split1_take_disjoint.csv	Take-disjoint: scenes shared across partitions, takes are not	9,715 / 585 / 1,794

Scene-03 is excluded from both splits: its indoor pool produces unreliable depth from the water surface, which was not annotated as a specular region. Per-scene composition is in the paper's supplementary material.

Versions

Version	Notes
v1.0-paper (tag)	Data as used for the WACV 2027 paper
main	v1.0-paper data plus metadata tidy-up: calibration files renamed left/right → az/elev, scene_id added to frame_index.csv, split files and this card added. No data files changed.

Citation

@inproceedings{deverobelfon2027crisp,
  title     = {{CRISP}: A Camera--Radar Dataset for Indoor {3D} Perception},
  author    = {Devero-Belfon, Alston Drew and Hanif, Md Abu and Santhalingam, Panneer Selvam},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027},
  note      = {To appear}
}

Acknowledgments

This research is supported by NSF grant 2348187.