Research · Founder work · 2024
Real-time character detection
Two YOLOv8 object-detection models trained on personal datasets of game footage, one for CS:GO and one for GTA Online's thermal view, running in real time on screen capture with a transparent overlay.
- 2
- custom-trained detection models
- 0.84
- precision on validation (recall 0.50)
- Live
- inference on screen capture
A computer-vision project covering the full training loop: collect and label data, train, read the metrics, change the dataset, train again. Games make a good test bed because the footage is easy to collect and the scenes change constantly.
What we built
- Personal datasets for two environments, iterated through several versions as the error analysis showed what the model was missing
- Training runs with YOLOv8 nano and small models, tracked through loss, precision, recall and mAP curves
- An inference script for local images and a real-time screen-capture loop with frame skipping and a confidence threshold, drawing results in a click-through overlay
- A conversion script for TensorRT to speed up inference on NVIDIA GPUs
Measured results
The best CS:GO run (YOLOv8s, 30 epochs at 640 px) reached 0.84 precision, 0.50 recall and 0.59 mAP@50 on its validation set. Precision is good, recall is the limit: the next step is more varied training data rather than a bigger model.
