CODEWITHSUNDEEP
tensorflowdatasetmask
K. Dmitry
K. Dmitry

Reputation: 11

Markup dataset for MASK-RCNN: only well-viewed objects?

I'm going to use mark-rcnn (based on tensorflow) to detect some cars and gasoline canisters. Images, what I have now, contains both cars and canisters. But I'm not shure about masks:

  1. Is it necessary to mark the object in image completely, or some of its parts are better not to allocate (if it concerns cars or gasoline canisters) - if there are several objects of the same type in the picture that need to be detected, but they stand one after another and partially overlap each other (for example, cars stand in the parking lot one after another)
  2. Do i have to mask each car, even if there is a small piece from the car, or do I have to mask with a certain visibility value (50%? 70%?), etc.? Or maybe better to avoid such images altogether and mark only those in which cars are fully visible?
  3. What else requirements are imposed to mask-rcnn architecture, where it would be possible to learn?
  4. What is the approximate configuration should be a computer for learning mask rcnn on images 1024x1024?

Upvotes: 1

Views: 119

Answers (1)

Adonis González
Adonis González

Reputation: 2068

1)I recommend that you just mark the object you need to detect, if it overlaps that part do not mark it, use polygon shape for your annotation!!!!!

2)The best would be to try to mark the whole car in each of your annotations, no matter how small or large it looks.

3)Python, use of libraries, I recommend you to use conda, and use tensorflow-gpu for your training it will be faster.

4)MaskRCNN has a function that manages that, I think you can find in utils.py

def resize_image():

this function resizes an image keeping the aspect ratio unchanged

Upvotes: 2

Related Questions