Computer VisionData-MiningDeep LearningMachine LearningProject

Developing a Self-driving Car – Day #3


Today I was able to test my new CNN architecture and the extended dataset. However preprocessing and augmenting the annotated images took much longer than I expected it would (It took around of 1h manual postprocessing and ~3h automatic post-/preprocessing). I started training my new NN architecture. My goal was 120.000 iterations. This takes about 3-5h on my GTX 1060.
But I wanted to see some first results before I would go to bed. That is why I aborted the training process after 18.000 iterations (At ~25.000 iterations the NN would have seen every image exactly once). I did not expect too great results, but the results after this short training time totaly overwhelmed me. I did not measure any kind of recall or precision. I only selected 40 random, not yet labeled, images and let my NN + Selective Search do the rest. The results look promising.
I share a few of these early results, so you are able to kind of get a feeling for what is working and where problems may remain (I censored persons and car plates to not run into privacy problems).

These results are quite promising for the short training time of 18k iterations. Tomorrow I will train the net for the full 120.000 iterations and test it again. I guess I need to do some finetuning, too. So my plan for tomorrow is to optimize my CNN. My currently used CNN looks like this:

Image created using: //ethereon.github.io/netscope/quickstart.html

I decided to do pooling not by using pooling layers but by using a kernel step of two for the first two layers.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.