The issue is that adding more training samples caused the SVM model to classify all images as non plate. Investigate the training script (likely in src/train) and verify the labels and preprocessing steps. Consider using a validation set to detect overfitting and check if new samples have different distribution than original ones.