Overview
The model(s) in this repository are trained with the dreamfactor/biggest-lego-dataset-600-parts from Kaggle and the Yolov7 training script.
Limitations
The zero-shot-1000-single-class.pt was trained in the training-zero-shot-1000-single-class.ipynb notebook with 1000 images and does not differentiate lego classes but only tries to predict Lego objects.
This can be easily reconfigured and retrained in the notebook, but the current implementation leads to many false positives on non-Lego objects and therefore can be improved
upon. Also, it could be worth investigating if the metrics improve with a bigger training dataset, as currently only 1000 images are being used (approx. 0.6% of the full
dataset).