Paired Image to Image translation Using U-NET generator, custom discriminator - PIX2PIX

Name: Santosh Vasa, Kishore Reddy Pagidi
Date: December 15, 2022
Class: CS 7180 - Advanced Perception

Operating System Used:

• Windows Sub-system for Linux - Windows 11. For training and development.

Link to the report

https://www.overleaf.com/read/qwghjktbxxwv

How to run?

• Install requirements.txt
  pip install -r requirements.txt
• Add the project folder to the PYTHONPATH:
  export PYTHONPATH=${PYTHONPATH}:/path_to_project_folder
• Update parameters.yaml with the data location.
• link to datasets :
  
  • https://www.kaggle.com/competitions/3d-object-detection-for-autonomous-vehicles/data
• Run python file train.py
  python3 train.py

RESULTS:

Areas where the network works well:

• Vehicles/classes near the ego vehicle
• Vehicles that are spaced apart.
• Non-overlapping line of sight vehicles.

Areas where the network does not work well:

• Objects occluded by other objects.
• Non square/rectangular segmentations.
• Post processing errors.
• Objects detected farther from the ego's position.

Example Inputs: BEV view afterlidar data projection

Outputs: In the order of (input, target and prediction)

Post Processing: