RetinaNet code example

@srihari-humbarwadi I’m following the keras code example on object detection, retina-net, and a few things are not clear to me.

1. anchor box: It uses an anchor box by assuming the target shape (xmin, ymin, width, height). Now, what if the target is (xmin, ymin, xmax, ymax)? Code.

class AnchorBox:
    """Generates anchor boxes.
    This class has operations to generate anchor boxes for feature maps at
    strides `[8, 16, 32, 64, 128]`. Where each anchor each box is of the
    format `[x, y, width, height]`.
....

What do I need to change of the above code? Direct Code link.

1. In data pipelines, what is the use of padding and the values, like below:

train_dataset = train_dataset.padded_batch(
    batch_size=batch_size, 
   padding_values=(0.0, 1e-8, -1), 
   drop_remainder=True
)

and why we use

train_dataset.apply(tf.data.experimental.ignore_errors())

[old-school-kid]

Do you mind giving some suggestions regarding anchor optimization for custom data set? Is it possible to auto-tune it by itself based on the custom data? (pointer)

As fas as I know this is SOTA mechanism when it comes to anchor boxes. And as always K-Means also produces quite satisfatory results. You can also try out IoU loss.

[srihari-humbarwadi]

I hope that the comments above do resolve your queries @Suzan009, and for

In data pipelines, what is the use of padding and the values, like below:

The model is trained on variable image sizes. Hence, to accommodate images of different sizes in a batch, we pad the images and labels.

• images are padded with 0s
• boxes are padded with a small float value 1e-8
• and class ids are padded with -1 (ignore label)