DEV Community

Aleksandr Gushchin
Aleksandr Gushchin

Posted on

July 12 - July 19 Weekly Status

This week I experimented with threshold according to one of my previous posts.

First thing I've done is to lower threshold itself. After a series of experiments I chose to decrease it by 35%.

You can see the results of lowering the theshold here:
Alt Text
To compare it with the previous results I post a picture from another blogpost:
Alt Text

The average F score on complete dataset was increased by ~0.096.

Also I tested different strategies to adapt threshold.

  • If max_interval option is specified and during max_interval frames metric below threshold algorithm chooses not the last frame of this series but the one with the biggest metric. You can see examples here: Before: Alt Text After: Alt Text

Max_interval option here is 500. You can see that before algorithm chose exactly 500 frame, but after it chose the one with the biggest metric.

Threshold does not take into account the following values of metrics:
Algorithm now don't mark the current frame as scene change if metric value of the next frame is bigger than the current one. It helps to prevent the algorithm to mark a series of consecutive frames as scene change (example below):
Alt Text

I didn't implement these changes in github repos yet and plan to do it early next week.

The most difficult remaining problem is the metric itself. I will take a closer look into it to see if there is a way to correct it a little bit. After it I will start to implement a new metric to compensate the weaknesses of this metric.

Top comments (0)