Week 12 - Long term prediction

less than 1 minute read

Long term prediction analysis

After obtaining a network that is capable of predicting with all motion types, I have carried out a new experiment in which the impact of increasing the time gap on network performance is analyzed. For this, I have generated 5 combined dynamics datasets that increase this value sequentially and I have made the prediction with our best network: LSTM-4. The results are shown below.

Long term prediction

In view of these results, it is clear that increase the time gap increase the error. However, the loss of prediction ability is always carried out within admissible thresholds: in a 640x480 image, for example, an average error of 14 pixels is obtained at 30 frames (1.7%) and 24 at 50 frames (3%).