ARCHITECTURE OF MobDepth:

• Input - Camera Frame (480,640,3) , Output - Depth Map (240,320,1)
• Design Choices:
  

  1. Real-Time Estimation - As the model needs to predict in real time, thus a light-weight architecture is required.
     As MobileNet uses least Number of Parameters without compromising on the Generalization, Thus it is used as Encoder.
     

     
     

  2. Generalization and High-Quality Depth Map - MobileNet is Pre-trained on ImageNet Dataset and including Skip-Connections make it more generalizable, removing the problem of Vanashing Gradient.

     SOME MORE RESULTS (on MobDepth(with Skip-Connections)):

     a) Fig-1
     b) Fig-2
     

AUGMENTATION OF SIMULATOR DATA (coppeliaSim):

• To integrate the component with robocomp, model should give high quality depth map on Simulator data as well, obained using coppeliaSim. Thus, NYU dataset is Augmented with Simulator Data.

• Process of Data Collection of Simulator Data can be found in Data-Collector

FINE-TUNING USING COLLECTED SIMULATOR DATA:

• The Pre-Trained Model (“MobDepth (with Skip-Connections)”) is trained further on collected simulator data to give high quality depth map in simulated world.

RESULTS BEFORE AND AFTER FINE-TUNING:

a) Fig-1

b) Fig-2