Unusual performance of the BPP of the fifth predicted frame in DCVC-RT

Dear DCVC Authors,

Thank you very much for releasing DCVC-RT and making the codebase publicly available. I am currently using DCVC-RT. I am interested in the trend and statistics of bpp within the GOP. I noticed that the bit rate per pixel (bpp) of the fifth predicted frame was quite unusual.

I combined 62 videos from MCL-JCV and UVG, including 720P and 1080P formats, totaling 310 GOPs. I set both reset_interval and GOP to 30, and QP to 0. I collected the results by sorting them by the sequence number within each GOP and conducted statistical analysis.

It was found that the mean value dropped rapidly and reached around 0.0025 in the 3rd to 4th predicted frames. The 6th predicted frame and thereafter remained stable at around 0.0025 and presented a wavy pattern, but the 5th predicted frame surged to 0.005.

Furthermore, the standard deviation started from the 2nd predicted frame, except for the 5th predicted frame, all were below 0.003 and remained stable at around 0.0015 and presented a wavy pattern. However, the 5th predicted frame surged to 0.008.

I know this is merely a statistical analysis based on my personal interest. I'm not sure about the reasons behind it, and I don't expect to receive any support or further feedback. As a student, I merely maintained my curiosity about the unusual phenomenon and wanted to present it to others.

If it is convenient, I would greatly appreciate it if you could share the information via email at: hushuy@mail.ustc.edu.cn

Thank you very much for your time and consideration.

Best regards.

Reproduce the reported bpp statistics with the 62 MCL-JCV and UVG videos, using reset_interval and GOP of 30 and QP 0. Compare bpp means and standard deviations by predicted-frame position, focusing on the fifth frame. Done means determining whether the fifth-frame surge is reproducible and identifying its cause.