Comparing cropping strategies

less than 1 minute read

Model Cropping Montmeló Simple circuit no red line Simple circuit no line no wall Simple circuit white road Simple circuit white road no line
Explicit -
PilotNet Normal
PilotNet NEW ✅✅ ✅(over the grass)
Deepest LSTM Normal
Deepest LSTM NEW ✅✅
Best PilotNet 3D Normal ✅(lowering rate)
Best PilotNet 3D NEW
Best Deepest LSTM 3D Normal ✅(lowering rate)
Best Deepest LSTM 3D NEW
Frankenstein (x3) Normal ✅(lowering rate) ✅(over the grass) ❌(last turns)
Frankenstein (x3) NEW ✅(over the grass)
Example image Normal
Montmelo normal
Simple no line normal
Simple no line no wall normal
Simple white normal
Simple white no line normal
Example image NEW
Montmelo new
Simple no line new
Simple no line no wall new
Simple white new
Simple white no line new

Best Frankenstein model: 20211229-161232_pilotnet_model_lstm_100_albumentations_depth_3_extreme_imgs_3dcnn_no_shuffle_cp.h5

Conclusions

Modifying the cropping of the input image in test time seems to help on the normal circuits (those close to the training images), whereas it doesn’t help when the circuit is modified.