Depth Estimation Prediction

The operation_mode top-level attribute specifies the operations or the service you want to execute.

In this tutorial, the operation_mode used is the prediction.

operation_mode: prediction

The general section and its attributes are shown below.

general:
  project_name: depth_estimation_fast_depth
  gpu_memory_limit: 16  # Maximum amount of GPU memory in GBytes that TensorFlow may use (an integer).

model: 
  model_path: ../../stm32ai-modelzoo/depth_estimation/fastdepth/Public_pretrainedmodel_public_dataset/nyu_depthv2/fastdepth_320/fastdepth_320_int8.tflite
  model_type: fast_depth
  

The model_path attribute is used to provide the path to the model file you want to use to run the operation mode you selected.

The gpu_memory_limit attribute sets an upper limit in GBytes on the amount of GPU memory TensorFlow may use. This is an optional attribute with no default value. If it is not present, memory usage is unlimited. If you have several GPUs, be aware that the limit is only set on logical gpu[0].

depth estimation requires images to be preprocessed by rescaling and resizing them before they can be used. This is specified in the 'preprocessing' section, which is mandatory in most operation modes. The 'preprocessing' section for this tutorial is shown below.

preprocessing:
   rescaling:
      scale: 1/127.5
      offset: -1
   resizing:
      interpolation: bilinear
      aspect_ratio: fit
   color_mode: rgb

Images are rescaled using the formula "Out = scale*In + offset". Pixel values of input images usually are integers in the interval [0, 255]. If you set scale to 1/255 and offset to 0, pixel values are rescaled to the interval [0.0, 1.0]. If you set scale to 1/127.5 and offset to -1, they are rescaled to the interval [-1.0, 1.0].

The resizing interpolation methods that are supported include 'bilinear', 'nearest', 'bicubic', 'area', 'lanczos3', 'lanczos5', 'gaussian', and 'mitchellcubic'. Refer to the TensorFlow documentation of the tf.image.resize function for more detail.

Please note that the 'fit' option is the only supported option for the aspect_ratio attribute. When using this option, the images will be resized to fit the target size. It is important to note that input images may be smaller or larger than the target size and will be distorted to some extent if their original aspect ratio is not the same as the resizing aspect ratio.

The color_mode attribute can be set to either "grayscale", "rgb", or "rgba".

In the 'dataset' section, users must provide the path to the directory containing the images to predict using the prediction_path attribute as well as the name of the dataset as shown bellow:

dataset: 
  dataset_name: nyu_depthv2
  prediction_path : ./datasets/nyu_depthv2/pred

The mlflow and hydra sections must always be present in the YAML configuration file. The hydra section can be used to specify the name of the directory where experiment directories are saved and/or the pattern used to name experiment directories. In the YAML code below, it is set to save the outputs as explained in the section visualize the chained services results:

hydra:
  run:
    dir: ./tf/src/experiments_outputs/${now:%Y_%m_%d_%H_%M_%S}

The mlflow section is used to specify the location and name of the directory where MLflow files are saved, as shown below:

mlflow:
  uri: ./tf/src/experiments_outputs/mlruns