I'm curious about the order of resizing and augmentations in the TensorFlow object detection API. For example, I'm using the config file ssd_mobilenet_v2_oid_v4.config. This uses fixed_shape_resizer and ssd_random_crop. So what is the interaction between these two modules?
Does the ssd_random_crop take a crop of the size defined in fixed_shape_resizer? If resizing happens first, then what size are the crops after resizing? And I assume they all need to be the same exact size in order to create proper batches?
Data augmentation happens before resizing. All preprocessing steps are specified in function transform_input_data in file inputs.py, this file contains functions like create_train_input_fn, create_eval_input_fn and create_predict_input_fn that will feed input image tensors to the model during training, evaluation and prediction. In create_train_input_fn, the following transform function is used.
def transform_input_data(tensor_dict,
model_preprocess_fn,
image_resizer_fn,
num_classes,
data_augmentation_fn=None,
merge_multiple_boxes=False,
retain_original_image=False,
use_multiclass_scores=False,
use_bfloat16=False):
"""A single function that is responsible for all input data transformations.
Data transformation functions are applied in the following order.
1. If key fields.InputDataFields.image_additional_channels is present in
tensor_dict, the additional channels will be merged into
fields.InputDataFields.image.
2. data_augmentation_fn (optional): applied on tensor_dict.
3. model_preprocess_fn: applied only on image tensor in tensor_dict.
4. image_resizer_fn: applied on original image and instance mask tensor in
tensor_dict.
5. one_hot_encoding: applied to classes tensor in tensor_dict.
6. merge_multiple_boxes (optional): when groundtruth boxes are exactly the
same they can be merged into a single box with an associated k-hot class
label.
Args:
tensor_dict: dictionary containing input tensors keyed by
fields.InputDataFields.
model_preprocess_fn: model's preprocess function to apply on image tensor.
This function must take in a 4-D float tensor and return a 4-D preprocess
float tensor and a tensor containing the true image shape.
image_resizer_fn: image resizer function to apply on groundtruth instance
`masks. This function must take a 3-D float tensor of an image and a 3-D
tensor of instance masks and return a resized version of these along with
the true shapes.
num_classes: number of max classes to one-hot (or k-hot) encode the class
labels.
data_augmentation_fn: (optional) data augmentation function to apply on
input `tensor_dict`.
merge_multiple_boxes: (optional) whether to merge multiple groundtruth boxes
and classes for a given image if the boxes are exactly the same.
retain_original_image: (optional) whether to retain original image in the
output dictionary.
use_multiclass_scores: whether to use multiclass scores as
class targets instead of one-hot encoding of `groundtruth_classes`.
use_bfloat16: (optional) a bool, whether to use bfloat16 in training.
Returns:
A dictionary keyed by fields.InputDataFields containing the tensors obtained
after applying all the transformations.
"""
The data augmentation is performed on step 2 (if there are any) and resizing is performed on step 4.