1. The first goal was to cluster 10.000 videos. For this purpose, we mapped the videos into 2D space by calculating their Google Si/Ti, determined, for example, in this paper. For calculating Google Si/Ti we used FFmpeg x264 codec with options [-qp 28 -b_qfactor 1 -i_qfactor 1].
2. Then, we performed a simple K-Means Clustering in this 2D space to divide the videos into 30 clusters.
3. Then, we measured the distances between centers of clusters and all the videos from the cluster. We chose 30 videos, closest to the centers of 30 clusters. These 30 videos included some hard-cases for deinterlacers, such as running letters, scene changes, motion, etc.
4. Another big part of dataset composing was to choose 60-frame cuts from these 30 big videos. To do that, we performed the same process, as in steps 1-3, but now the goal was to choose 30 cuts from about 15.000 cuts. Again, we calculated Google Si/Ti for each 60-frames cut and clustered them into 30 clusters. 2 clusters contained only black and only white cuts respectfully, so we marked them as "trash" clusters.
5. Then, we selected top-5 cuts, closest to the center of each non-"trash" cluster.
6. The final step was to manually choose the best from top-5 cuts for each non-"trash" cluster.
7. We don't count PSNR on black frames. They were made for deinterlacers that use motion estimation (ME). Because of the black frames, ME-deinterlacer detects zero motion between two sequences and, therefore, doesn't consider motion from the previous sequence while processing the current one.

The main criteria are that the PSNR between GT fields and the same fields in the deinterlaced video must be equal to infinity. To control this, we make Top-Fields and Bottom-Fields plot on every second frame of GT video and deinterlaced video.

Here is the sample plot for Bob-Weave Deinterlacer. This deinterlacer passed the validation.

PSNR between Bottom-Fields of every deinterlaced frame and GT frame is equal to infinity, so we substitute infinity by zero. This means, that the bottom field exactly matches the corresponding field in the GT sequence.

Another sample plot, for MSU Deinterlacer, which had problems with colorspace.

As we can see here, Bottom-Field PSNR is about 100, but not infinity. In such cases, we make the following steps:

1. Convert GT data to the colorspace of the deinterlacer.
2. If we don't know (can't guess) which colorspace the deinterlacer uses, we form the LookUp-Table from the fields of the deinterlacer output and GT data, which must be equal to each. Then, we use this LookUp-Table to map GT to the colorspace of the deinterlacer.
3. In some cases, it is impossible to precisely determine LookUp-Table because the mapping is neither injective nor surjective. In such hard cases, we just choose a GT sample with the highest PSNR from the previous 2 steps.

Also, we provide MSU Video Quality Measurement Tool (VQMT) PSNR visualization of the deinterlacer output. As the last step of the validation process, we check, that the VQMT PSNR visualization is striped. It should be so because even/odds rows must exactly match the corresponding GT rows.

Here is the example of correct VQMT PSNR visualization output:

And, finally, let us take a closer look at it:

As we can see, the output is stripped, so this means that the deinterlacer passed the last validation step.