Can untrained labels still show up in detection?

Hi all, I have one simple question as I am curious if people would consider that a bug or if it is some obscure ML implementation by design with Tensorflow: Can a model “detect” an object label in which no photo data was given to it to make the association?

for example, I start a project where I create many labels, all the labels I will possibly need for my project, but I only use a few labels so far in annotating the pictures I have now, and then export the entire project into a tfrecord file for training that will contain all my labels and the photos that have the annotations consisting of only some of those labels (all labels not used). I do some transfer learning to train a model for object detection, convert the model using tensorflow.js converter, load it up and run an inference. I get a detection for a label I provided no data for to the model. Is this a bug (either program or user), or is it by design by how Tensorflow and machine learning stuff work?

Hi @CharlesVelazquez, Generally this behavior is known as Zero-Short-learning where a model can recognize a category of objects in images without ever having seen a photo of that type of object. This happens when a model is trained on a sufficient amount of data. If the features learned during training are relevant to the unseen class, the model might be able to recognize it to some extent. Thank You.

@Kiran_Sai_Ramineni That’s. . . . scary, and also means more work for me unfortunately, but oh well. Thanks so much for the insight.