Can computer vision help us understand imperfect mimicry?

Nobody has claimed this yet.

Hi everyone!

I work with wasp mimic moths and have a small preliminary dataset (≈150) that I will keep building over the next year with sympatric wasps and moths I collected around Panama. The group of moths I work with has an amazing range of mimic fidelity, from species almost unrecognizable to some species that hardly resemble any known hymenoptera but that have been regarded as mimics in the community for their transparent wings, flight behavior, and diurnal activity. I study wing transparency in this group with different focuses, and in one of them it’s the community assemblages. Although these are really charismatic species, we lack information about the models they are mimicking, something essential for a basic understanding of their mimicry phenotypes. Most studies on wasp mimicry have focused on fly mimics that don’t exhibit the amount of phenotypic diversity that this group of moths does. My aim is to understand how they are mimicking and which traits are being selected by predators (which traits define the mimicry phenotypes). My approach is to use computer vision to cluster these sympatric insects to understand these questions by identifying who is in the clusters and what the machine used to cluster them.

I plan to use t-SNE in a similar way as described in a recent paper [https://royalsocietypublishing.org/doi/10.1098/rspb.2019.1501],with the difference that I will not use a body template, as I believe body shape is biologically relevant. I am an evolutionary biologist with basic knowledge of computer vision and ML, so I would like to take this opportunity to discuss this idea with the rest of the team and get some insight on how to approach it better, keeping in mind that my dataset is still in progress.

No repository files, tests, or implementation entry point are identified. Start by reviewing the proposed t-SNE approach and the linked paper, then clarify the intended contribution, data format, and evaluation criteria with the project team. Done cannot be defined from the current issue because it is an open research discussion rather than a scoped implementation task.

Written by the indexing model from the issue text.