jerredchen00/image-as-an-imu-finetuning

Image as an IMU: Real-world Finetuning Dataset

Official real-world finetuning dataset from Image as an IMU: Estimating Camera Motion from a Single Motion-Blurred Image (ICCV 2025 Oral).

[arXiv] [Webpage] [GitHub]

[PIXL, University of Oxford](https://pixl.cs.ox.ac.uk/)

Jerred Chen, Ronald Clark

Dataset Details

This dataset consists of 32 sequences of real-world motion-blurred videos in various indoor scenes, captured using the iPhone 13 camera.

dataset_train_real-world.csv and dataset_val_real-world.csv are the CSV files used for training/validating the model in the paper. These can be directly plugged into the provided dataloader in the GitHub.

The CSVs provide the following:

• —blurred: the relative path to the (resized 320x224) motion-blurred RGB image
• —ts1,ts2: the frame timestamps between the previous RGB and next RGB image
• —fx,fy,cx,cy: the scaled camera intrinsics, corresponding to the 320x224 image
• —bRaqx,bRaqy,bRaqz,bRaqw: body-frame rotational velocity, parameterized as a quaternion
• —btax,btay,bta_z: body-frame translational velocity
• —exposure: exposure time at the given image
• —sequence: the sequence name

Sequence Details

Each sequence consists of the following:

sequence1/
├─ blurry_frames_320x224
│  ├─ XXXXXX.jpg
│  └─ ...
├─ confidence
│  ├─ XXXXXX.png
│  └─ ...
├─ depth
│  ├─ XXXXXX.png
│  └─ ...
├─ rgb
│  ├─ XXXXXX.jpg
│  └─ ...
├─ rgb_320x224
│  ├─ XXXXXX.jpg
│  └─ ...
└─ blurred_frames_320x224.csv
└─ camera_matrix.csv
└─ camera_matrix_320x224.csv
└─ imu.csv
└─ odometry.csv
└─ velocities.csv

Sequences were recorded using the StrayScanner app, slightly modified to also obtain the exposure time from ARKit. confidence, depth, rgb, camera_matrix.csv, imu.csv, and odometry.csv are the original outputs from StrayScanner.

We provide the following data in addition to the StrayScanner outputs:

• —rgb_320x224 are the resized recorded RGB images
• —blurry_frames_320x224 are the identified frames with more extensive blur using FFT as described in Liu et al.
• —camera_matrix_320x224.csv are the corresponding scaled camera intrinsics
• —velocities.csv consist of the translational velocities computed from ARKit poses in odometry.csv and the rotational velocities directly from the gyroscope.

Of course, the RGB images/camera intrinsics can be resized/scaled online during training; we provide this to maintain consistency with our own training.

Since the ARKit computed poses can have very large errors, dataset_train_real-world.csv consists of manually filtered samples without large outlier pose estimates.