0

I would like to train an Object detector using Mobilenet SSD Model on a custom dataset.

Looking at the config file of Mobilenet, there is a block called image_resizer{} which I think the default is 300x300, but my available images is in 224x224.

Is it okay for me to carry on the training without altering the config file or do I really need to change it to 224x224 to match my images?

Found the config file here.

https://github.com/tensorflow/models/blob/d6d0868209833e014074d6cb4f32558e7acf2a6d/research/object_detection/samples/configs/ssd_mobilenet_v1_pets.config#L43

kjell324
  • 1
  • 2

1 Answers1

0

Image resizer is a fixed resize - so if you stage with the default (300x300) - the images will be resized to be 300x300 before they passed to the network.

One thing to keep in mind is the default number of branches you have in this model - which is 6 - when you start from 300x300 input the final branch will be 1x1, In case you decide to change the input size, you may also need to update the number of branches (a.k.a num_layers in the config) - because there is only room for 5 branches (in case you don't change their arch.) - you actually get to 1x1 after 5 branches.
To be more clear here are the branches you will have starting from 224x224

  1. branch 0 - 1/16 of the image - 14x14
  2. branch 1 - 1/32 of the image - 7x7
  3. branch 2 - 1/32 of the image - 4x4
  4. branch 3 - 1/32 of the image - 2x2
  5. branch 4 - 1/32 of the image - 1x1
Tamir Tapuhi
  • 406
  • 3
  • 8