The datagen command CLI is highly confusing, regarding the image quantities, as it means a tile quantity for training/validation on the first hand, and a raw image quantity for testing on the other hand.
The reason of this choice was justified as follows:
we pick a high number of random bbox in the training/validation images in order to build a robust tile dataset ;
whilst we only want to cover up the whole image in the testing case: we generate tiles following a regular grid defined by the tile size.
One should clarify the CLI, by harmonizing the -t, -v and -T definition: these arguments should refer to the number of raw images (like in the testing case). As a consequence, one could add an additional argument (e.g.--nb-tiles-per-image) to define the targetted amount of tiles obtained from a single image, for the training/validation cases.
The
datagencommand CLI is highly confusing, regarding the image quantities, as it means a tile quantity for training/validation on the first hand, and a raw image quantity for testing on the other hand.The reason of this choice was justified as follows:
One should clarify the CLI, by harmonizing the
-t,-vand-Tdefinition: these arguments should refer to the number of raw images (like in the testing case). As a consequence, one could add an additional argument (e.g.--nb-tiles-per-image) to define the targetted amount of tiles obtained from a single image, for the training/validation cases.Linked to #146