How to collect my own dataset?

[ShengtongZhu]

What a owesome project!! Thanks a lot~
You gave me a lot of advice as a beginner.
I really want to know how to collect my own stop sign and right_curve datset.I mean how many photos should I take for per sign ?
I can never thank you enough for your help.

[ShengtongZhu]

Excuse me~

[jawilk]

Apologies for the late reply, I hope it's still helpful.

You can see the final dataset I used for training here:
https://github.com/jawilk/Self-Driving-RC-Car-Payment/tree/master/program/traffic_sign_detection/data
The quantities are as follows:

• Background: 4001 images
• 50 km/h: 1529 images
• Priority: 1678 images
• Stop: 1333 images
• Dangerous Curve Right: 362 images
• No passing: 935 images

Unfortunately this task is very time intense.
(i) Started with this dataset http://benchmark.ini.rub.de/?section=gtsrb&subsection=dataset and looked at every image if it's similar to mine.
(ii) Took around 50 images of each class with my phone and cut them on my PC.
(iii) Augmented the images i already got (i.e. rotate/skew/mirror etc.)
(iv) After I got a roughly working network, I put 4 signs of the same type in a row and let the car drive along collecting images, extract region proposals and make predictions. After that I added them to the right label folder. This saved a lot of time compared to taking pictures with my phone and cutting by hand and also assures that the training images are almost the same as the real images the network has to predict later on (didn't pay much attention to overfitting here, since the detection was supposed to only work in that specific environment (yet)).

Let me know if something is unclear or you have further questions.