Non-iterative optimization of pseudo-labeling thresholds for training object detection models from multiple datasets