Deep neural network for traffic sign recognition systems: An analysis of spatial transformers and stochastic optimisation methods
Pytorch Implementation of Deep neural network for traffic sign recognition systems: An analysis of spatial transformers and stochastic optimisation methods
- ÁlvaroArcos-GarcíaJuan A.Álvarez-GarcíaLuis M.Soria-Morillo*
Deep neural network for traffic sign recognition systems: An analysisof spatial transformers and stochastic optimisation methods
https://reader.elsevier.com/reader/sd/pii/S0893608018300054?token=0656FA2921430AA401BA73A6990A187F32A6FBDD12EAA2FC87FD556B3CDDF6DA8D5BE54F230A979E57369C48AB081452
LCN Implementation is taken from https://github.com/dibyadas/Visualize-Normalizations
- ASGD Works best among all optimizers for me for Learning Rate : 10^-2
- Class imbalance is removed prior to training by duplicating the data
- Learning Rate Decay worsenes the performance
- Data Augmentation, in general, decreases performance although Spatial Transformer model ensures augmentation isn't a problem.
- Architecture is changed slighlty from the original set of layers
- Currently Gaussian filter is kept constant for LCN, where as ideally it should be chosed at random during run-time
Neural Net gives output of 6 neurons necessary for Affine transformation (translation, cropping, rotation, scaling, and skewing) and uses grid generator and sampling as inbuilt Pytorch commands
- Max Jaderberg Karen Simonyan Andrew Zisserman Koray Kavukcuoglu
Spatial Transformer Networks (https://arxiv.org/pdf/1506.02025.pdf )