Generaly, Neural Network based algorithms should be more robust compared to hard coded algorithms. A similar case should be expected for our Visual Control algorithm as well. In this post, we study the performance of the brains by changing the camera position.
Camera Offset
The images obtained by different camera positions are:
| Original | Left | Right | Rotate Up | Rotate Down |
|---|---|---|---|---|
 |
 |
 |
 |
 |
PilotNet
Following are the graphs
Position Deviation Error
Comparision of Position Deviation Error
Position Deviation (MAE)
Comparision of Position Deviation (MAE)
Percentage Completed
Comparision of Percentage Completed
Lap Seconds
Comparision of Lap Seconds
Average Speed
Comparision of Average Speed
Deepest LSTM Tiny PilotNet
Following are the graphs
Position Deviation Error
Comparision of Position Deviation Error
Position Deviation (MAE)
Comparision of Position Deviation (MAE)
Percentage Completed
Comparision of Percentage Completed
Lap Seconds
Comparision of Lap Seconds
Average Speed
Comparision of Average Speed
Explicit Method
Following are the graphs
Position Deviation Error
Comparision of Position Deviation Error
Position Deviation (MAE)
Comparision of Position Deviation (MAE)
Percentage Completed
Comparision of Percentage Completed
Lap Seconds
Comparision of Lap Seconds
Average Speed
Comparision of Average Speed
Observations
-
Car having the camera rotated up, never completes the circuit.
-
For the Deep Learning brains, the rotate up and rotate down, both of them do not complete the circuit. However, the rotate down completes the lap for Explicit Brain.
-
The Right and Left camera offsets are able to complete the circuits, but, have a higher error compared to the original camera position.
Conclusions
It was expected for the PilotNet and DeepestLSTM networks to perform well for the Left, Right and Rotated Down camera offsets. For Left and Right, this was well observed, but for Rotate Down the brains did not perform the same. This may be due to the croping that is done before processing the image. Due to that, a particular field of view of the image may be blocked that is helpful for the network to work.