Open tyiannak opened 3 years ago
Yes, of course! I will discuss with @electrasif for the experiments.
Can I make the experiments with the old version of analyze_visual.py? The new version with object detection wich using PyTorch is very very slow and there are many experiments.
no i would prefer to use the current version. if it is too slow u can temporarily cache the features (and not commit the code) for each video file
The first phase of the experiments was completed. Our goal was to find out how each variable affects precision and recall. The method we followed to achieve this was to keep two of the three variables constant and change values in the remaining variable. Below we suggest which range of variables we consider appropriate to perform grid search.
if ((mag_mu> [0.04,0.05,0.06,0.08,0.1]) and (gray_diff > [0.002,0.02,0.04,0.05,0.06]) and (f_diff[-1] >[0.22,0.55,0.75,0.95,1,5]))
Are the above values right to proceed with the grid search process? Below is the spreadsheet with the experiments. (https://docs.google.com/spreadsheets/d/1JgvNMInXPCZjImzHFCqa1mJlwntHmxZ2DjScwJtimQ8/edit?usp=sharing)
@apoman38 @electrasif can you split the experiments and select the optimal set of parameters? Use master branch and the shot_evaluation script for different thresholds in the shot_change. In particular, create a set of values for the thresholds in line 411 and report the precision and recall in a spreadsheet. Lets try 30 or 40 combinations of threshold values and select the best. You can split experiments in half. Can you have this until tuesday @electrasif @apoman38 ?