You are working to create an object detection system, like the ones described in the lectures, to locate cats in a room. To have more data with which to train, you search on the internet and find a large number of cat photos.
Which of the following is true about the system?
should I add internet pictures to train/test/dev set?
I think I should, even these pictures have not been taken in the room.
From what I believe will happen, I would not add too many of these outside pictures if the model-serving scenario is always inside a room. For instance, if your model runs on a kitchen surveillance camera, it is good to have cats in a room setting in the pictures.
Maybe from the internet also, look for pictures of cat in a room setting. Too many pictures of cats in the wild will skew the weights learned by the model. I would keep <1% pictures of cats in the wild.