Clarification regarding limit on number of parameters

[sama...@bu.edu]

Hi all,

In the rules, we mentioned that participants are allowed to use a model with at most 300 million parameters and that 2 forward passes using a 150 million parameter model for a single image would be counted as 2 x 150 million parameters for this limit.

Here we wish to clarify that we shall enforce this rule on a per-pixel basis for segmentation models, meaning, the number of parameters is the maximum value of forward passes times number of model parameters for any given pixel. 

Hence, if you choose to do inference over multiple non-overlapping windows for a single image using a model with 150 million parameters, although you may run a forward pass multiple times, any individual pixel of the image is processed by a model of at most 150 million parameters, which keeps it under the limit of 300 million.  

This clarification has been added to the rules on the challenge webpage : https://ai.bu.edu/visda-2022/

Best,

VisDA-22 challenge organizers