[objection_detection] Data augmentation input format scaling confusion

[apatsekin]

System information

• What is the top-level directory of the model you are using: not relevant
• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): No
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 18.04
• TensorFlow installed from (source or binary): binary
• TensorFlow version (use command below): 1.15.0-rc3-22-g590d6ee
• Bazel version (if compiling from source): not relevant
• CUDA/cuDNN version: 10.0
• GPU model and memory: not relevant
• Exact command to reproduce: not relevant

Describe the problem

Problem concerns data augmentation/preprocessing in Object Detection API.

1. In inputs.py:

def augment_input_data(tensor_dict, data_augmentation_options):
    tensor_dict[fields.InputDataFields.image] = tf.expand_dims(
      tf.cast(tensor_dict[fields.InputDataFields.image], dtype=tf.float32), 0)

We see that image is being casted to float32, as initially it's uint8 (I double-checked it).
2. Then few lines further image (as a part of tensor_dict is passed to preprocessor.preprocess):

 tensor_dict = preprocessor.preprocess(
      tensor_dict, data_augmentation_options,

3. Which according to header expects pixel values in [0,1] range (and all augmentation functions as well), but gets [0.0,255.0] since from uint8 only casting happened.

tensor_dict: dictionary that contains images, boxes, and can contain other
                 things as well.
                 images-> rank 4 float32 tensor contains
                          1 image -> [1, height, width, 3].
                          with pixel values varying between [0, 1]

Am I missing something or [0,1] in comments to every function in preprocessor.py is not relevant and it actually deals with [0.0,255.0]

[netanel-s]

Pretty sure it's just a mistake, and all preprocessing functions receives and acts on [0,255].
The normalization from [0,255] to [-1,1] happens in the preprocess of each feature_extractor.
Did you see any function which actually expects [0,1]?