When I started Snap2Console, the challenge was simple but fun:
Can we take a random game cover photo and recognize which console it belongs to—and even identify the exact game?
I collected Japanese cover art, trained an EfficientNet_B2 model for console recognition, and paired it with k-NN galleries for game recognition.
Console recognition: 95–98% accuracy (depending on setup).
Game recognition: ~20% accuracy, limited by having only one image per game.
Best performance came from large datasets like PlayStation and Saturn, while smaller classes (like Mega Drive) showed lower recall.
Overall, the project showed that:
Neural nets do really well at console classification.
Game recognition is much harder without a richer dataset (multiple images per title).
Augmentation alone isn’t enough—you need more diverse images.
You can check out the code on GitHub.
What’s Next?
This was a great project to dive deeper into computer vision pipelines, but now I’m moving on to a new challenge—most likely something related to large language models (LLMs).
I want to explore how LLMs can be applied beyond text generation, for example in interactive systems or AI-powered applications that combine language and other modalities.
Stay tuned—new project coming soon!
I want to explore how LLMs can be applied beyond text generation, for example in interactive systems or AI-powered applications that combine language and other modalities.
Stay tuned—new project coming soon
Top comments (0)