German Traffic Signs Recognition Benchmark
Introduction
The German Traffic Sign Benchmark is a multi-class, single-image classification challenge held at the International Joint Conference on Neural Networks (IJCNN) 2011. The competition is designed to allow for participation without special domain knowledge.
The benchmark has the following properties: Single-image, multi-class classification problem More than 40 classes More than 50,000 images in total Large, lifelike database
Kaggle competition on the The German Traffic Sign Recognition Benchmark [http://benchmark.ini.rub.de/?section=gtsrb&subsection=news] was conducted as a part of Computer Vision class.
Experiments
- 3- layer Convolutional Network as base test
- ResNet 18
- VGG
- Spatial Transformer Network modules with modified layers
- Spatial Transformer Network modules with ResNet18
Preprocessing
- Rescale- Rescaled the images to 48x48 and 64x64. When rescaled the images to 48X48, the results were better than 32x32.
- Transform- Tried to use rotate Transform for ResNet18 but didn’t get good results.
Conclusion
The images in dataset have severe flaws such as blur, shadow, obstructions, bad lightening. There are distortions like rotation, translation, etc. These are difficult for human to recognize.
Achieved best accuracy of 99.06% with Spatial Transform Netwrok variant on the final test set portion. Spatial Transformer networks are good for classification tasks with lot of distortion in data. As, STNs take care of color and spatial variance and avoid augmentation, with learned color and spatial variance the model can adapt to image’s characteristics.