I am trying to train a model with tf2 to find ambiguous objects such as small damages or small deformations in vehicles. I have tried several models but I have to lower the threshold to a lot so that it detects something … do you advise me some special model or some configuration for the pipeline.config? Thank you very much
From the my understanding, if the classes are very similar, you might need a very good dataset with lot’s of examples.
How many images/boxes are you using?
about 8700 images and about 6000 boxes … but the problem is that it is very difficult for the models (I have tried several) to detect these (both damage and deformation) types of objects … I need to lower the threshold of .30 (by default ) to find something … any ideas? Thank you so much
do you think the resolution of the input image could be too low? to a point where the model wouldn’t have enough information to make a decision?
I don’t think so because the naked eye is visible … all the images are 800x600 and most of the boxes are visible to the naked eye … I don’t think the data set is the problem, I have tried efficient models (d2 and d3) to lower the learning rate because if I did not start training, it gave me a nan in losses … the truth is that I am a bit lost and I don’t know which parts of the pipeline to modify or if I should look for another model … many thanks for your interest and help.
Hi,
You say to the naked eye the difference is visible.
However, before your image is fed to the Neural Network it is likely reduced in size to fit the input layer.
Check the input size of your model, then take a couple of images, reduce them to that size and check again to see whether you can still determine the boxes by eye. You could be surprised.
Regards,
Chris
1 Like