Mobilenet SSD Input Image Size

Question

I would like to train a Mobilenet SSD Model on a custom dataset.

I have looked into the workflow of retraining a model and noticed the image_resizer{} block in the config file:

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

Does the aspect ratio here have to be 1:1 like 300x300 or can I specify a custom ratio?

All my dataset images are 960x256 - so could I just input this size for height and width? Or do I need to resize all the images to have an aspect ratio of 1:1?

Solution

Choose the height and width, in the model file (as per your link), to be the shape of the input image at which you want your model to train and operate. The model will resize input images to the specified size, if it has to.

So this could be the size of your input images (if your hardware can train and operate a model at that size):

image_resizer {
    fixed_shape_resizer {
        height: 256
        width: 960
    }
}

The choice will depend on the size of the training images and the resources required to train (and use) that size of model.

I typically use 512x288 as this size model runs happily on a Raspberry Pi. I prepare training images, at a variety of scales, at exactly this size. So the image resizer does no work during training.

For inference, I input images at 1920x1080, so the image resizer scales them to 512x288 before they pass into the Mobilenet, maintaining the aspect ratio.

However, the aspect ratio is not important in my domain since such distortions occur naturally.

So yes, just use your training image dimensions.

Comments

[Tesla.]

Hi, thanks for your answer. I looked into the keep_aspect_ratio_resizer and in this post here: stackoverflow.com/questions/48145456/… they said that you can't use it with SSD because it ends in fully connected layers. The problem I see with setting it to 300x300 and letting the network resize it is that the aspect ratio is not 1:1 and the images will be distorted in the input layer and the objects might not be detectable, especially smaller ones.

[Nahiyan]

I think you should edit your answer as it may mislead people to use non-square images with Mobilenet SSD model.