GermanTrafficSignDetection
Intro
This project builds a Deep learning model that localizes (with a bounding box) and identifies German traffic signs.
The algorithm is an adaptation of Fastai's Lesson 9 to the problem of Traffic Sign Detection (i.e. bounding box regression + identification) on the German Traffic Sign Detection dataset
Results Summary
Below are results on 16 validation set images
For a better view of each results, go to the end of this Readme.
Conclusion
Although the results are far from satisfactory, we were able to reach encouraging results with only ~500 training examples. Here are a list of things we could do to improve these results:
- Add more data
- Use a multi-pyramidal architecture, such as RetinaNet
- Conserve higher-res images for training (in order to avoid memory issues with my GPU, I had to downscale images to 448x448)
Appendix 1: zoom-in on each result
