I’m a bit confused about the test split and the loss of actual bounding boxes to evaluate a trained neural network.
Whats the test split for if I cannot produce bounding boxes for object classes to evaluate the performance of my neural net with scenes the network has not seen during training?
Hi. Please read https://towardsdatascience.com/train-validation-and-test-sets-72cb40cba9e7
You train on the training set and validate on the val set. Any ablation studies in your paper need to be on the val set, but it is easy for a user to fake good results, as they have the ground-truth annotations. Therefore a separate test set exists for which the ground-truth annotations are only stored on our evaluation server, such that the user cannot cheat (unless he reannotates the dataset).
Thanks so much, this totally explains everything. I wasn’t aware of the competitions as I’m not researching in OD or tracking.