It seems odd that only one object of each class is tracked in a video segment. I'd imagine that it would be a limitation for algorithms that generate bounding boxes for each class, as they might be penalized for correctly finding the non-tracked instances of a given class. Is only tracking a single instance of a class standard for this kind of dataset?

I can't say what is the best source of noise to experience this, but I have heard what sounded like my brother's voice on aeroplanes several times.

Could you suggest some good papers/articles on the topic?

Any info on the network itself (number of layers, input resolution, pooling, dropout etc)?

Stock Cafenet, trained from scratch.