Traffic Sign Recognition: a performance comparison on challenging weather conditions between CNNs with spatial transformers and ResNet50
The task is the classification of a traffic sign among 43 possible classes over real images of traffic signs captured in Germany.
The classification has been done implementing the state-of-the-art models for this task (according to GTSRB Benchamark) and exploring the potentiality of ResNet50 as backbone for transfer learning.
The comparison between the models is presented in our report.
- Traffic Sign Recognition: Classify the traffic signs from images.
- Convolutational Neural Networks: Utilize convolutional neural networks for feature extraction.
- Convolutional Neural Network with Spatial Transformers: Incorporate spatial transformers to handle geometric transformations and improve accuracy over CNN architecture.
- Transfer learning using ResNet50: Training of the last fully connected layers of the backbone keeping the weights for all the other layers locked.
- GTSRB Dataset: German Traffic Sign Recognition Benchmark (GTSRB)
- Description:
- Weather augmented datasets (Weather50-Weather100): Adding weather augmentations images to GTSRB dataset with a respectively fixed probability of the augmentation to be applied of 0.5 and 1.
- Models can be found here: https://drive.google.com/drive/folders/1P0Vv7XsjSRJ90Za2QqmW0gQHPsd9ucO1?usp=drive_link
Classification examples of a convolutional neural network with spatial transformers on weather aug- mented dataset.
The effects starting from the left are: rain, shadow, sun flare, fog. On the top row the predicted label with its confi- dence, while on the second row the ground truth class.
- Python 3.11
- PyTorch 2.3.0
- Other libraries listed in
requirements.txt