I have managed to set Digits and Caffe up & some train models.
But I have difficulties to get correct results from Predictions.
For example, if I try search which road sign the image is…
It works in some way, as if I have analyse between different shapes of road signs, for example, “Stop” road sign and basic “speed” road sign is easy to analyse. and Digits do it very well. Because those road signs are very different in basic shape.
However, the problem comes if I try analyse road signs that have same round basic shape, but different internal image, for example.
No parking image - https://upload.wikimedia.org/wikipedia/commons/9/9c/Road_Sign_No_Parking.jpg
No motorcycles - https://upload.wikimedia.org/wikipedia/commons/thumb/2/2b/Singapore_Road_Signs_-Restrictive_Sign-No_motorcycles.svg/553px-Singapore_Road_Signs-Restrictive_Sign-_No_motorcycles.svg.png
I also use few other images with similar differences.
Digits and Caffe seems to have difficulties to recognize which road sign is in the picture. It can also give complete false results.
Should I change some setups to get more correct results?
- I have tried to decrease model’s “Base Learning Rate”, but no help.
Is there any CUDA library that could do that?
Is there any CUDA library that could find road signs from the image, no matter which is their location in the image?