Datasets:

OmniInstrument
/

SLAM_project

language:
  - en
license: mit
license_link: LICENSE
pretty_name: omnislamproject
task_categories:
  - robotics
  - depth-estimation
  - keypoint-detection
configs:
  - config_name: default
    data_files:
      - split: stereo
        path: data/stereo-*
      - split: stereoinertial
        path: data/stereoinertial-*
      - split: vio
        path: data/vio-*
dataset_info:
  features:
    - name: image_left
      dtype: image
    - name: image_right
      dtype: image
    - name: timestamp
      dtype: float64
    - name: gyro
      list: float32
      length: 3
    - name: accel
      list: float32
      length: 3
    - name: sync_dt
      list: float32
      length: 2
    - name: position
      list: float32
      length: 3
    - name: orientation
      list: float32
      length: 4
  splits:
    - name: stereo
      num_bytes: 381314848
      num_examples: 100
    - name: stereoinertial
      num_bytes: 381134104
      num_examples: 100
    - name: vio
      num_bytes: 370323077
      num_examples: 100
  download_size: 1132892975
  dataset_size: 1132772029

Omni Instrument SLAM Project Dataset

The Omni Instrument SLAM Project Dataset is a compact robotics dataset designed for evaluating stereo, visual-inertial, and visual-inertial odometry (VIO) pipelines.

It provides:

Stereo Image Pairs
Inertial measurements (IMU)
Ground-truth 6 DoF pose (for VIO)
Raw ROS 1 and ROS 2 recordings

Overview

The dataset is structured into three splits:

Split	Description
`stereo`	Stereo-only (IMU stationary)
`stereoinertial`	Stereo + IMU
`vio`	Stereo + IMU + ground-truth pose

All splits share the same schema, enabling consistent downstream pipelines.

Data Collection Protocol

AprilTag grid settings (used for both Stereo and Stereo-Inertial calibration sequences):

Setting	Value
Tag size	20 mm
Tag spacing	6 mm
Tag family/dictionary	`tag36h11` (6x6)

Stereo (Calibration - Static Sensor)
- Setup: robot stationary, camera + IMU fixed, AprilTag grid moves
- Used for: stereo calibration (intrinsics/extrinsics)
Stereo-Inertial (Calibration - Moving Sensor)
- Setup: robot moves, camera + IMU move together, AprilTag grid stationary
- Used for: camera-IMU extrinsics + sync checks
VIO (Operational SLAM Sequence)
- Setup: robot moves in a normal environment (no calibration targets)
- Logged: stereo images, IMU, ground-truth odometry
- Used for: VIO/SLAM evaluation

Data Format

Each example follows the same top-level schema. Some fields are split-dependent:

stereo: images + timestamp (IMU is stationary; pose not provided)
stereoinertial: adds IMU (gyro, accel) and time offsets (sync_dt)
vio: adds ground-truth pose (position, orientation)

Example record:

{
  "image_left": Image,
  "image_right": Image,
  "timestamp": float,

  "gyro": [wx, wy, wz],
  "accel": [ax, ay, az],

  "sync_dt": [dt_right, dt_imu],

  "position": [x, y, z],
  "orientation": [qx, qy, qz, qw]
}

Notes:

gyro is in rad/s and accel is in m/s^2.
sync_dt = [dt_right, dt_imu] are time offsets (in seconds) relative to timestamp (left image):
- dt_right = abs(t_right - t_left)
- dt_imu = abs(t_imu - t_left)

Sampling Methodology

Each split contains 100 randomly sampled, synchronized frames:

Uniform sampling across the trajectory
Start/end trimmed to remove initialization artifacts

Synchronization constraints:

abs(t_left - t_right) <= 5 ms
abs(t_left - t_imu) <= 5 ms
abs(t_left - t_odom) <= 5 ms (VIO only)

Missing Data Handling

For splits without ground truth (stereo, stereoinertial):

position    = [inf, inf, inf]
orientation = [inf, inf, inf, inf]

ROS Topics

Recordings:

ROS 1 bags: stereo, stereointertial
ROS 2 MCAP: vio

ROS 1 (Calibration)

/stereo/left/color/image_raw
/stereo/right/color/image_raw
/imu/data

ROS 2 (VIO)

/stereo/left/color/image_raw
/stereo/right/color/image_raw
/imu/data
/ground_truth/odom
/tf

Example Usage

from datasets import load_dataset
import numpy as np

ds = load_dataset("OmniInstrument/SLAM_project", split="vio")
sample = ds[0]

img_l = sample["image_left"]
img_r = sample["image_right"]

gyro = sample["gyro"]
accel = sample["accel"]

pos = sample["position"]
quat = sample["orientation"]

if not np.isinf(np.asarray(pos)).any():
    print("Ground truth available")

License

This software and dataset are released under the MIT License.