Designed by MrPioSmasher for MrPio smashing. 🦆
A computer vision pipeline that performs real-time player detection and automated targeting in Blockstorm. Built entirely from scratch as a deep learning research and experimentation project.
This is an improved version of this earlier attempt.
| Functionality | Status |
|---|---|
| Image segmentation | ✔️ |
| Bounding box detection | ✔️ |
| Head detection | ✔️ |
| Mouse movement | ✔️ |
| Testing | 🔳 |
The project features two modes: image segmentation or bounding box detection. In both cases, the model was trained from scratch.
A U-Net architecture trained to segment the game screen and detect opponent pixels. Since it performs pixel-level classification, it is more accurate, but slower.
A simplified, custom YOLO-like architecture trained to detect bounding boxes around players. It is lightweight and optimized for high frame-rate.
In both cases, a simple algorithm is used to detect the head's likely position, starting from the model outputs.
