Future approaches - HelmchenLabSoftware/contrast-marker-video-tracker GitHub Wiki

Clearly, there are exceptions for any simple rule we can devise to identify the spots. Therefore, it is necessary to do a composite criterion, which would include a compromise between several different approaches.

1. Correlation analysis

One of the original ideas was to correlate regions of nearby frames, and thus deduce the possible future position of the region based on its past position, or vice-versa. While alone this method lacks predictive power, due to ROI changing shape or disappearing altogether, it can be used as an aid.

2. Dynamics analysis

Possibly, additional information can be extracted from the difference between next and previous frame, or the acceleration or other metrics. As of now, the frame rate is too slow for this to be a good estimator, because between two frames ROI can move from invisible to completely visible, and move distances significantly larger than the ROI size.

3. Supervised Learning

A completely different approach can be taken by applying machine learning. A Deep NN can be taught on movies already identified by hand. Ultimately, such algorithm may be able to identify markings without the need for additional contrast enhancement and any user input except visual validation, which would significantly accelerate experiment and analysis.