The pictures in this blogpost will depict two charts each - they differ only in the gray vertical lines. In the upper chart the gray lines represent ground truth, in the lower chart the result of the algorithm.
Blue line represent values of algorithm's metric on frames, orange - threshold.
My first observation is that the threshold is often too high. It can illustrated in this picture:
The second observation is that parameters (min interval and max interval) are often bad for the algorithm (especially the max parameter):
You can see that algorithm marked exactly 500th frame as scene change although there are frames before it with higher metric.
The third observation is that often the algorithm does not work well because of the threshold, but because of the metric:
The fourth observation - the threshold in general does not behave optimally, it does not take into account the following values of metrics, because of this a lot of consecutive frames can be marked as scene changes:
You can see that around 2000th frame the metric consistently grows, followed by a delayed growth of the threshold. Because of this, about 100 consecutive frames are marked as scene changes. Here is closer look on it:
Summary:
- Threshold often is way too high for single spikes of metric values
- Max interval parameter can be improved, it better to mark previous frames with higher metric values
- Threshold does not take into account the following values of metrics
- Metric can be wrong (usually either the metric works well on the whole video or poorly on the whole video)
Top comments (0)