I have a model (a convolutional neural network) that has already been trained on set of training data for human pose estimation. It works very well for simple cases, but I want to make it more robust to human truncation and occlusion. For this I prepared a different set of data that covers such "difficult" situations. I intend on taking the current weights and biases as initial values for the model and train it using gradient descent and back-propagation on this new set of data. As I lack experience in deep learning, before I go on and apply this, is this "correct" from a scientific point of view? should I change something before (such as the loss function...)?
If you learn separably on two different data sets you can expect second data set training will blur results from first training. You can avoid this by stop second training early however estimate in what point is very difficult task and still you won't have guarantee that second training do not obliterated first training results. So finally you can expect that after two training your model will have higher quality response to data similar to second data set, and lower quality to first data set (comparing to model learnt only on first data set).
I strongly recommend you to merge two data sets if its possible (you can use already learnt network to learn on merged set), or prepare additional flow in higher layers of your network.