Welcome to the ultimate showdown of Computer Vision vs. Pikachu! As I zigzag through the electrifying path of Computer Vision, I've decided to tackle the most shocking challenge first - detecting Pikachu in the wild (and in pictures, because let's be honest, the wild is hard to come by).
To develop and refine computer vision models capable of spotting the electrifying presence of Pikachu. Why Pikachu, you ask? Because if you can detect an electric-type Pokémon adept at camouflaging itself in various environments (and occasionally in Ash Ketchum's arms), what can't you detect?
Since there are a lot of mature frameworks available for computer vision tasks by great teams such as Ultralytics, Meta, Google, etc, this repository acts as a central place to learn how to work with these with a common goal of training them to work with my custom dataset.
- Phase 1: Learning how to not electrocute myself while plugging in my GPU.
- Phase 2: Collecting a dataset of Pikachu images and labelling them.
- Phase 3: Training models to detect Pikachus.
- Final Phase: World domination (or at least, winning a local Pokémon Go contest).
| Framework | Task | Architecture |
|---|---|---|
| Ultralytics YOLOv8 (WIP) | Object Detection | YOLO |
| Ultralytics YOLOv8 (WIP) | Semantic Segmentation | YOLO |
| Tensorflow Object Detection API (WIP) | Object Detection | Faster RCNN |
| Tensorflow Object Detection API (WIP) | Semantic Segmentation | Mask RCNN |
| Meta Detectron2 (WIP) | Lorem | Ipsum |
| Meta SAM (WIP) | Lorem | Ipsum |
- Add the code for ultralytics YOLO experiments
- Compare different model sizes of YOLO
- Add the code for tensorflow object detection API experiments
- Explore Meta's Segment Anything Model (SAM)
- Explore Meta's Detectron2
- Multi class examples (Where's My Pikachu and Ash?)
- Contribute Pikachu Images: Got a Pikachu photo? Is it high quality, low quality, or just right for training a neural network? Head over to the dataset folder to read the process that I followed for labeling images. You can help increase the pikachus in the dataset.
- Code Contributions: If you're skilled in the art of computer vision or just want to learn alongside me, your pull requests are more than welcome.
- Suggestions: Are there some frameworks/tools that you use at your workplace? Create an issue about the tools you'd like me to train on custom data.
This project is for educational purposes and a personal challenge on my journey to deepen my computer vision knowledge. It's also a testament to my love for Pokémon. Please use it responsibly and don't try to catch actual Pikachu with it. They don't like it.