Can we Train Deep Learning Model with inclusion of negative samples?

Question

Hi everyone,

Background

I have trained a model for detecting cars from orthophoto involving only land area.

However, it also wrongly detects ships as "cars" when I applied it to coastal area.

I have limited time to run variation of "training".

I am using mask RCNN in my training.

Q1. Do "Train Deep Learning Model" GP use unlabeled tiles as background?

Q2. So will it help reducing treating ships as cars if I adding unlabeled tiles with ships?

Q3. I see Output No Feature Tiles option in Export Training Data For Deep Learning.

Will it be useful if I use it to generate background image? I am afraid it will much increase my training              time.

Thank everyone for helping with these concept.

2.jpg

1.jpg

ChristianHarrington_2-1618260243101.png

MariusN_0-1717759413522.png

Felix10546 · Answer

Thank TimG for answering the questions.

I tried a export with no Feature Tiles option. However, they are seems ignored during the training process.

So I think the answer for Q1 and Q2 is no?

Q3. Thank you for your suggestion. I tried and it solved my problem

TimG · Answer

Hello

Q1: Pull some label tiles from the label directory into ArcGIS Pro. Click on the non car areas - if it says 0 its background. Unlabeled areas/tiles area usually background.

Q2: Not 100% on this but yes I think it should.

Q3: I will increase training time, assuming your tile count goes up.  Also when exporting tiles of ships, you will probably need to set a region of interest, so an area mask. To prevent getting huge amounts of data/tiles

Regards

Tim