kocurvik/nico_stereo_data

NICO stereo vs. RGB-D: data

Data for "Built-In Stereo vs. RGB-D: Practical Depth and 6D Object Pose Estimation on the NICO Humanoid" (Palider and Kocur, ICETA 2026). Code: https://github.com/kocurvik/nico_stereo.

The same scenes are recorded with the built-in wide-angle stereo pair of the humanoid robot NICO, an Intel RealSense D435i and a Stereolabs ZED M. The repository holds the raw recordings and the evaluation data (calibrations, noise model, per-image depth metrics, FoundationPose 6D poses) behind the paper's tables and figures.

Download

Every path is relative to the root of the code repository, so downloading into it gives the layout the code expects:

git clone https://github.com/kocurvik/nico_stereo && cd nico_stereo
hf download kocurvik/nico_stereo_data --repo-type dataset --local-dir .            # everything, 11.3 GB

To fetch only a part, add --include patterns from the table, e.g. the files needed to check every number in the paper (paper_scripts/verify_data.py):

hf download kocurvik/nico_stereo_data --repo-type dataset --local-dir . \
    --include "out/out_24042026/cameras_parameters/*" "out/out_24042026/cameras_statistic_model/*" \
              "out/out_24042026/depth_comparison/*" "out/out_24042026/inference_time_stats.csv" \
              "out/out_24042026/pose_estimation/3D_models/*" "out/out_24042026/pose_estimation/masks/*" \
              "out/out_24042026/pose_estimation/results/*" "out/out_24042026/pose_estimation/results_check/*" \
              "out/out_24042026/depth_estimation/*"

Contents

D = datasets/dataset_24042026, O = out/out_24042026.

part	--include	size	contents	needed for
out core	O/cameras_parameters/* O/cameras_statistic_model/* O/depth_comparison/* O/inference_time_stats.csv O/pose_estimation/{3D_models,masks,results,results_check}/*	0.3 GB	calibrations, noise model, per-image depth metrics, CAD models, masks and the FoundationPose poses of all arms	every scoring script
depth samples	O/depth_estimation/*	0.03 GB	run_stats.json of all 13 depth sources and two depth maps each (frames 0 and 3). Not the full set of depth maps, see below	verify_data.py, Fig. 3
pose inputs	O/pose_estimation/depth_*/* O/pose_estimation/undistorted_images_NICO/*	0.9 GB	depth of the six pose scenes from every depth source (16-bit PNG, mm) and the undistorted NICO images	re-running FoundationPose
stereo 4K depth	D/stereo_4k_depth/*	5.5 GB	the 215 evaluation frames: 4K stereo pair, RealSense and ZED M RGB and depth	re-running depth networks, re-scoring from raw
downstream task	D/downstream_task/*	1.5 GB	the six pose scenes (001, 002, 004, 005, 006, 009)	re-running the pose pipeline from raw
calibration	D/stereo_4k_calibration/* D/stereo_4k_relative_pose/* D/calibration_ZED/* D/calibration_Realsense/* D/distance_validation/*	1.5 GB	calibration and validation images	re-running calibration
camera stats model	D/camera_stats_model/*	1.6 GB	repeated static scenes behind the RGB-D noise model	re-running the noise model

The predicted depth maps of the 13 depth sources (about 190 MB each) are not included; the inference scripts in the code repository re-create them. What is included is everything computed from them: O/depth_comparison/zed/metrics_cauchy/ holds the per-image and summary metrics, and the depth maps used for the pose scenes are in the pose inputs. Depth maps are metric depth in metres, one <frame>_depth.npy per frame at 640x360. FoundationPose results are 4x4 poses in O/pose_estimation/results/<arm>/<scene>/<object>/ob_in_cam/<frame>.txt.

License

CC BY 4.0. Please cite the paper when you use the data.