For the training set, noise and saturation pixels are synthesized on-the-fly in GyroDVD_model.py. Please refer to RSBlurPipeline.
For the validation and test sets, noise and saturation pixels have already been applied and are included in the released dataset.
- GyroVD_Syn_train.tar.gz : 50,500 frames for training.
- GyroVD_Syn_train_tau.tar.gz : pre-computed tau. (only required for training.)
- GyroVD_Syn_val.tar.gz : 5,000 frames for validation.
- GyroVD_Syn_val_tau.tar.gz : pre-computed tau for validation set. (only required for training.)
- GyroVD_Syn_test.tar.gz : 7,700 frames for evaluation.
- GyroVD_Syn_test_tau.tar.gz : pre-computed tau for test set. (only required for training.)
- GyroVD_Real.tar.gz: 10,000 real-world frames for evaluation.
To facilitate future research, we also provide accelerometer and magnetometer data.
The GyroVD-Syn dataset
# GyroVD_Syn_train.tar.gz
GyroVD_Syn_train
├── 0413/VID_20250412_101526
│ ├── blur # blurred frames
│ │ ├── 000001_160466752524897_160466785733026_160466786578871_avg09_blur.png
│ │ ├── 000001_(exp_start_1th)_(exp_start_9th)_(exp_end)_avg09_blur.png
│ │ ...
│ ├── gt # gt frames
│ │ ├── 000001_160466752524897_160466785733026_160466786578871_avg09_gt.png
│ │ ...
│ ├── meta_info
│ │ ├── gyro.csv # gyroscope data
│ │ ├── accel.csv # accelerometer data
│ │ ├── magnetic.csv # magnetometer data
│ │ ├── cam_metainfo.csv # camera meta data
│ ├── rot # pre-computed rotation matrices
│ ├── sat_mask # saturation mask for RSBlur pipeline
├── 0413/VID_20250412_101616
│ ...
...
The GyroVD-Real dataset
# GyroVD_Real.tar.gz
GyroVD-Real
├── 1029_GP/VID_20251029_013931
│ ├── blur # blurred frames
│ │ ├── 000001_193795012257259_193795012257259_193795111113483_blur.jpg
│ │ ...
│ ├── meta_info
│ │ ├── gyro.csv # gyroscope data
│ │ ├── accel.csv # accelerometer data
│ │ ├── magnetic.csv # magnetometer data
│ │ ├── cam_metainfo.csv # camera meta data
│ ...
...