Object Detection on small test images with medium train image bounding boxes

[Zeakles]

I want to train a model with tensorflow faster rcnn that can detect animals from a "far" distance where the objects are relativly small (example: https://cdn1.spiegel.de/images/image-830326-breitwandaufmacher-bfrb-830326.jpg ).

Much of my training data is was captured closer (example: https://www.welt.de/img/reise/nah/mobile139849438/0572506457-ci102l-w1024/Schafe-in-bei-Clifden-irische-Kleinstad.jpg ) How can my model detect smaller objects better that look pretty similar to my training data but are smaller?

I already tried to augment my Data in the pipeline (faster rcnn config). This improved my results lot but its still not perfect.

  data_augmentation_options {
    random_horizontal_flip {
  }
  }
  data_augmentation_options {
     random_pad_image {
  }
  }

I also tried to lower the scales of my anchors to detect the smaller objects in my dataset better, but this had a bad impact on the model

I also tried to add the same picture in different aspect ratios to the tfrecords, which had till now a bad impact on the model. What to you think i can change to get better results with my model?

[redeipirati]

Hi @Zeakles,

I want to train a model with tensorflow faster rcnn that can detect animals from a "far" distance where the objects are relativly small ... Much of my training data is was captured closer.

This is something that it's not good since you are basically changing the data distribution. You should have a dataset that is fully representative of the task you want to solve, in this case, your training data should only contain images of animal from the distance.

How can my model detect smaller objects better that look pretty similar to my training data but are smaller?

I've never tried this before, maybe these two resources can give you a good starting point:
- How do you do object detection using CNNs on small objects like ping pong balls?
- R2-CNN: Fast Tiny Object Detection in Large-scale Remote Sensing Images
- Finding Tiny Faces

Hope that helps.