Dataset scales when training with custom data about os2d

[aosokin]

Hi Simon,

The main goal of resizing images is to align object and images-to-detect with model expectations. Object images are internally resized to have maximum size of 240, and this number is baked in the architecture - the number of parameters of the Transformation network (more details here). Os2dHead internally resizes (here) feature maps to match the expected size.

So to get the best performance on your own dataset in the eval mode, you need to use an image pyramid such that at least one of its levels has objects-to-detect of size approximately 240. For training, you need to select the sampling of patches to match the corresponding sizes. All of these can potentially be set by approximately computing dataset_scale for your dataset. We did exactly this for all datasets we touched in the paper (e.g., grozi).

Hope this helps!

Best,
Anton

[chasb799]

Thank you for your helpful respone!