Vineeth Bhat

(Written earlier; GitHub permalink)

Find week 1 at Trying to Teach Freshmen Robotics 1.

1. Weeks 5 to 8
   1. Stuff that I sent over
   2. PyTorch
   3. Robotic Vision with OpenCV
2. Final Things

Weeks 5 to 8

Stuff that I sent over

1. PyTorch
2. Robotic Vision (with OpenCV)

Evidently, progress in the stuff I could teach was slow owing to academics (ironic, isn’t it?). Regardless, they are still interested in RRC and a couple of them got part-time internships in a robotics startup (Yay!).

PyTorch

Not just the basics but also dive into critical aspects that will be essential for applying neural networks to robotics problems.

1. Getting Started with PyTorch:

• Resource: Deep Learning with PyTorch: A 60 Minute Blitz
  Familiarize yourself with the core components of PyTorch.

1. Data Loaders and Datasets:

• PyTorch’s torch.utils.data.Dataset and DataLoader classes are designed to simplify data handling.
• Key Aspects:
  • Custom Datasets: How to structure and preprocess your robotics data for training.
  • Batching & Shuffling: Efficiently loading data during training to improve performance.
  • Parallel Data Loading: Leveraging multi-threading to accelerate your training pipeline.
• Resource: The Data Loading Tutorial

1. Optimization Theory:

• Understanding the mathematics and intuition behind optimization is crucial.
• Focus Areas:
  • Gradient Descent and Variants: Learn why and how different optimizers work (SGD, Adam, etc.).
  • Tuning Hyperparameters: Experiment with learning rates, momentum, and other parameters through torch.optim.
• Resource: The PyTorch Optimizers Documentation is a solid starting point.
• Other stuff:
  • You have done quite a bit of this in the mathy bits earlier and will redo it in the mobile robotics course. Take a look again at this in for the time being: Intro to Optimization

Robotic Vision with OpenCV

1. Introduction to OpenCV:

• Resource: The OpenCV-Python Tutorials for an overview.

1. Multi-view Geometry:

• A super important thing in robotic vision is the ability to perceive depth and spatial relationships using multiple images.
• Focus Areas:
  • Camera Calibration: Learn to determine intrinsic and extrinsic camera parameters to correct lens distortions.
  • Resource: The Camera Calibration Tutorial guides you through this process.
  • Epipolar Geometry & Stereo Vision: Understand how to estimate depth and recover 3D information from two or more camera views.
  • Resource: Multi-view Geometry is perhaps the best book available. There are a few copies in the library as well.

1. Robust Vision Pipelines:

• Key Concepts:
  • Feature Detection: Experiment with detectors such as SIFT, SURF, or ORB to recognize and match key points in images.
  • Real-time Processing: Build pipelines that can handle continuous video streams and process images on the fly.

Final Things

Once you’re done. Start with reading about SLAM pipelines.