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
bRa_qx,bRa_qy,bRa_qz,bRa_qw: body-frame rotational velocity, parameterized as a quaternion
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.