Deeper Exercise: YoloV5 Road Dataset

[henrylao]

Overview

(group) Exercise Deeper: CHOOSE one of the following (with instructors permission) depending on your Project 3: (50 points) DUE Nov. 19 at 11pm (you have the option to turn this in during fall recess by Nov.26 11pm) -- I WANT MOST TEAMS TO DO one of the first 2 options --only if you convince me will I allow you to do training on cloud. Create Video of your exercise running and MAKE sure to go through the code AND show the results and any output/files/code created and testing results. Turn in to BB->Exercises->Exercise Deeper. Ideally your video will be on youtube and you are turning in a URL. It is good to make videos like this and you can even "show" potential employers.

• Source: http://borg.csueastbay.edu/~grewe/CS663/outline.htm (11/09/21)

Deliverables

• Implementation of frontend code
• Implementation or usage of the backend code (make sure to apply comments)
• YouTube demo the model on Android posted w/ walkthrough of code
  • Show results (Tensorboard outputs from training, testing, validating, etc)
• Documentation (this is essentially done from Exercise Explore) further work can include

  • PyTorch models to clarify official vs unofficial architectures: https://pytorch.org/vision/stable/models.html
    - Professor Grewe was pretty adamant that YoloV5 from ultralytics was not official (she's correct however ultralytics's yolov5 appears to be quite well developed with an extremely active community of devs behind it): https://pytorch.org/hub/ultralytics_yolov5/

  • Some simple gotchas Android's pickiness with file naming evident in using periods <str>.<str>.ptl (PyTorch Obj Detect Droid Demo) as well as underscores<str>_<str>.pdl (PyTorch Img Seg. Droid Demo)

  • Can also include how to compile YoloV5 for Torch Mobile ie follow the YoloV5 Repository instructions instead of following TorchHub's instructions NOTE pretty positive I missed some important/simple things since I was operating in a rush for the 11/09 due date

  • Include how models can be configured in PyTorch -- specifically in the case of YoloV5, we can configure through the YAML file. It's possible that we can modify using the code directly; however, since we're using the CLI implementation this is likely unnecessary

Relevant Files

Google Colab Notebooks and Resources can be found at:

• Notebook: https://colab.research.google.com/drive/14jHT1BtXV94CDZ1UFDKlIDPu7dp7Icie?usp=sharing
• Directory Configuration for Training: https://drive.google.com/drive/folders/1EFTBSuMhQq9AtznS0T_i_dXdWCDzB3r0
• Relevant Training Results, Graphs, etc: https://drive.google.com/drive/folders/1QMBbVlnxUJ1zrhethni2ymu9MeF5R9qA?usp=sharing

Our PyTorch Mobile Android:

• https://github.com/henrylao/pytorch-yoloV5-roadside-detection

Demo of the current Droid App (UPDATED 11/09/21):

• YoloV5 Demo: https://drive.google.com/file/u/2/d/1ccBP_zwbvTWwPT48tHewKsVoJnd4s_67/view?usp=sharing