So in the lecture, Prof. Ng mentions that if we have multiple classes to predict, then we may apply Non-max suppression individually to each class. However, in the programming assignment (Car detection with YOLO), we have multiple classes to predict, but we use the function " tf.image.non_max_suppression()", which applies non-max suppression just once and does not run through each class individually.
Now as far as I can understand, using NMS individually for each class may result in a higher accuracy, but it might cost higher computation and time to run. However, not using NMS individually for each class may result in a lesser accuracy, but it might save on computation and time to run. So, I can see that there is a tradeoff in following both approaches, and the decision may very on the goals and evaluation metrics.
This is what comes to mind. Please let me know if I’m understanding it correctly, and also your thoughts on this, and what could be the pros and cons of both approaches. Thanks!