1 of 16

ME5990

�

Project 2 Kickoff

2 of 16

Outline

  • Dataset and Project
  • Strengthened Slides as a Report
  • Demonstration and GUI
  • Some Tips

​

3 of 16

Learning is beyond “Image recognition”

  • Deep learning is considered “end to end”
  • Deep learning model is a function that takes in the “data” and throw out the “label”.
  • A good model means higher accuracy on telling the correct “label”

​

​

Model

Input data

Label

Image Recognition Task: Cat

Image Caption Task: Cat with sunglasses

Object Tracking Task: Nose (greenbox), mouth (redbox)

4 of 16

Many learning tasks you can do

  • Computer vision:
    • Image recognition
    • Object Tracking
    • Image/video captioning
    • Semantics Segmentation
    • 3D reconstruction
  • Robotics
    • Control
    • Odometry/Pose estimation
    • Path Planning
    • Autonomous navigation
    • SLAM
  • Natural Language Processing
  • Finance
  • Sociology…

5 of 16

Open-source dataset

  • Really many
    • Kaggle: Basically everything, usually light weight
    • KITTI/ CityScape : Autonomous navigation
    • Yahoo Finance Scrapping

​

6 of 16

Selected Example Dataset

​

​

​

  • Input: RGB image
  • Label: Segmented Image with the same size

​

Model

Label

Input

7 of 16

Selected Example Dataset

​

xmin

xmax

ymin

ymax

class_id

2

135

130

230

1

Model

car

8 of 16

Selected Example Dataset

9 of 16

Selected Example Dataset

​

 

Model

10 of 16

Project Description

  • Select a topic of your own interest (No image recognition anymore)
    • Must use supervised learning method
  • Search the available dataset
    • You can collect data yourself, if you have the time…
  • Train at least 2 models for the dataset
    • ChatGPT permitted
    • Import pre-trained model: not permitted
    • Your model must have major differences as existing models
  • Conduct evaluation on your model
    • Possibly: compare your result with benchmark results

11 of 16

Deliverable

  • Presentation at the last lecture
  • Enhanced presentation slides as the final report

​

12 of 16

Discussion on Selection

  • Select Project and Dataset
    • Not the more complicated the better
    • Choose the one that intuitively make sense
    • Understand what is the input, and what is the label
  • Dataset
    • Some datasets are too big to use for us
    • Beware: some datasets need train-test split (KITTI), some not.
    • Kaggle offers reasonable ones
      • Select ones with high usability
    • If you choose to make your own…
      • Label may be a big issue, unless you have methods to auto label (For example: https://github.com/dmariaa/ets2-data-capture )

13 of 16

Model

  • You must understand each layer of your model
  • You shall understand the input and output tensor dimension of each layer
  • You shall be able to estimate the parameters in the model
  • ChatGPT is permitted
  • Pre-trained models are not allowed
  • Other people’s model can be referred; but shall not be used as it is.
  • Try at least 2, but you probably will try tons

14 of 16

Training

  • Report the training parameters
    • Entropy method
    • Optimizer
    • Learning rate
    • Epoch, batch
    • Etc.
  • Report the hardware you use
  • Report your training time

15 of 16

Evaluation

  • Big Headache… Different problems have different rubrics
  • Refer to peer submissions, what are their rubric?
  • Compare your result with others
  • You can create your own rubric
    • Make sure you can clearly explain it

16 of 16

Timeline

  • December 4: Presentation (10 min per group)
  • December 11: Submit your final slide as the report