Labeling policy for airplane detecting YOLO

Question

I am training my YOLO to detect airplanes and drones. in some of the pictures it is impossible to detect that the object is indeed an airplane, and it even looks like a drone (pictures are taken from very far away), but I know from the context that it does. should I still label it as an airplane?

Answer 1

The correct answer might require a system level look on the problem.

For instance, if you have the budget to run 2 models one way to tackle it would be:

1. Train a Detector of a flying object (Airplane or a Drone).
2. Train on the bounding boxes of (1) a classifier.

The real question to answer here is on your system, what is more important, detect either or make a classification error?
This question should be the design goal of the solution.

Specifically, it is best to label by the prior knowledge.
Yet in cases the object is small, and for your system requirements it does not matter if the classification is Drone or an Airplane, you may use some label smoothing so both classes have relatively high score in the Ground Truth.

Answer 2

It should be our hope that ML can find patterns in otherwise noisy or imperfect data. Unfortunately, some training examples may only confuse an algorithm but with any luck, dicey examples simply have a low confidence.

It is a practitioners job to give true labels. The dataset may later need to be curated to get optimal performance. But a good heuristic is likely to start with a fully and correctly annotated dataset and go from there.

Answer 3

Split into two problems:

1. Binary classification whether there is an object is flying.
2. Binary classification whether it's airplane or not.

Both of them using separate model instance.

Answer 4

Yes, you should still label it as airplane. The airplane that may resemble drone features for noise to a human eye, needs not be labeled as drone as you have other context to know it is an airplane.