Synthetic Results: City Scene

Back to the index

PSNR of each frame in the input sequence and in the results of the 3 methods in compaison. Our method performs best in terms of preserving texture details.

Frame 19

Please move your mouse cursor over the small images to toggle between the results.

All the three denoising algorithms perform fairly well on the city sequence. Our algorithm is better than the other two at preserving some of the finer details, most noticeably the texture in the street and trees and the edges of the windows. Difference images with respect to the ground truth are shown below for the full frame.

Frame 46

All three denoising algorithms perform well. When toggling between each result and the ground truth, our result is noticeably closer to the ground truth near the red brick building and trees on the right. Difference images are shown below for the full frame.

Alternative flow methods

The above experiments use our flow as input for both our and Liu and Freeman's denoising methods. Now, we further evaluate the two denoising methods by using Liu's flow estimation and the ground truth flow as input.

Flow method from [12]

This is the flow method used by Liu and Freeman [13]. The flow is very inaccurate for frames with large displacement such as frame 19. Results of running their denoising algorithm and our denoising algorithm on the inaccurate flow are shown.

Liu and Freeman's method produces significant artifacts when the flow is bad. Our algorithm includes a flow reliability check and falls back on spatial denoising, making it more robust to flow inaccuracy.

Ground truth flow

Our algorithm is better able to preserve details by using the reliable flows and information from any higher quality frames nearby. Liu and Freeman's method results in more smoothing since it still includes spatial denoising.

Back to the index