Dependency requirements

This project requires the following dependencies:

• Python 3.8
• opencv-contrib-python==4.4.0.46
• timm
• torch==1.9.0
• torchvision==0.10.0
• pandas
• requests
• tqdm
• matplotlib
• seaborn

To install the dependencies via pip:

pip install -r requirements.txt

DepthEstimation

This project uses the MiDaS Depth Estimation model, trained on 10 datasets (ReDWeb, DIML, Movies, MegaDepth, WSVD, TartanAir, HRWSI, ApolloScape, BlendedMVS, IRS), which predicts the depth of each pixel in the image. We employ the DPT_Hybrid model weights.

Initialization

Initialization example:

from DepthEstimation.core.models.depth_model import DepthEstimator

model_settings = {
   'device': 'cpu' # Device string used for pytorch (options: 'cpu'| 'gpu')
}
system_config = {'debug':
                    {
                    'debug_mode': True  # Bool Flag
                    },
                 'depth_estimation':
                    {'stages': 3, # Number of distance levels
                     'grid_size': 5, # Square size of objects center used for depth estimation
                     'model_type': 'DPT_Hybrid' # Depth estimation model's name
                     }
                 }

depth_estimator = DepthEstimator(
                model_settings=model_settings,
                system_config=system_config['depth_estimation'],
                debug_mode=system_config['debug']['debug_mode']
            )

Input/ Output

This addon needs to be executed after inference as it relies on the information of the detected bounding boxes:

• Input:

  coords = addon_object.inference.boxes
  scores = addon_object.inference.scores
  classes = addon_object.inference.classes
  
  # estimate depth of every pixel of an inout image
  depth_img = self._estimate_depth(addon_object.frame)

• Output

  # Ordered bounding boxes, scores and classes 
  addon_object.inference.boxes = sorted_coords
  addon_object.inference.scores = sorted_scores
  addon_object.inference.classes = sorted_classes
  # Distance IDs for each corresponding bounding box, the IDs correspond to the configured number of stages and grib size
  addon_object.inference.extra['distance_ids'] = stage_distance_ids