Notes.org

Notes

Note

In all the experiments below, the numbers reported from [Jiang et. al.] are based on a model trained on 1014 youtube videos in addition to the Adobe240FPS dataset. Since these curated videos are not publicly available, my models were trained using the following publicly available data sources:

1. NFS dataset - http://ci2cv.net/nfs/index.html - this data is 240 FPS
2. Vimeo dataset - http://toflow.csail.mit.edu/ - this data is only 30 FPS
3. Adobe dataset - https://www.cs.ubc.ca/labs/imager/tr/2017/DeepVideoDeblurring/

Evaluation on Adobe240FPS dataset

Source	PSNR↑	IE↓	SSIM↑
SuperSloMo [Jiang et. al]	31.19	8.30	0.918
SuperSloMo [this implementation]	32.039	7.257	0.927
SuperSloMo-R [this implementation]	34.124	5.896	0.951

Evaluation on Slowflow dataset

Source	PSNR↑	IE↓	SSIM↑
SuperSloMo [Jiang et. al]	34.19	6.14	0.924
SuperSloMo [this implementation]	36.311	4.322	0.938
SuperSloMo-R [this implementation]	37.988	3.568	0.955

Evaluation on Vimeo dataset

Source	PSNR↑	IE↓	SSIM↑
SuperSloMo [this implementation]	34.736	4.813	0.951
SuperSloMo-R [this implementation]	35.554	4.488	0.956

Evaluation on Sintel high frame rate dataset

Source	PSNR↑	IE↓	SSIM↑
SuperSloMo [Jiang et. al]	32.38	5.42	0.927
SuperSloMo [this implementation]	31.799	5.859	0.921
SuperSloMo-R [this implementation]	32.714	5.241	0.932

Evaluation on Sintel optical flow benchmark (training set + final pass)

Source	EPE↓
DSTFlow	7.82
Unflow-CSS	7.91
MFO2	6.01
SuperSloMo [this implementation]	7.36
SuperSloMo-R [this implementation]	6.97