| Our Coded Sampling Input | Our Output | Ground Truth |
| Global 4x Exposure | Global 1x Exposure | Denoised Global 1x Exposure |
The global exposure videos have 1/4 the high-speed frame rate to match the sampling rate of our coded input.
Press the button in the bottom right corner of the video to view fullscreen.
Our result is more accurate on the surface of the chessboard than the 4x exposure video, which is partially saturated. Some saturation artifacts remain on the hand in our result because the ADMM method has not converged yet after 50 iterations. The images below show the effect of more or fewer iterations. Regions without saturation converge much faster than those with saturation.
|
|
|
|
|
| 25 iterations | 50 iterations | 75 iterations | 100 iterations | Ground truth |
Our result has some trouble with the fast and complex motion of the dice. The constant exposure videos retain the detail of the dots better; however, the denoised video loses some of this detail due to over-smoothing. Our method performs better in the dark regions than the 1x exposure plus denoising, which doesn't recover any of the wood grain texture on the table or side of the chessboard.