search

uas-at-ucla / suas-23-24

Software/Vision operations for UAS@UCLA, 2023-2024. For a list of tasks, visit https://github.com/orgs/uas-at-ucla/projects/1
3 stars 0 forks source link

Generate emergent target training set #29

Open nathanchan631 opened 1 year ago

kushhansagarwal commented 1 year ago

Will these DALL-E images work fine? I mean they kinda look AI generated but I don't think the detector would/should fail at classifying them.

Image Image Image Image

EW0824 commented 1 year ago

Umm, interesting.

The only thing though, the resolution might be too high and we want the training data to resemble real data as much as possible. Definitely play around with DALL-E/midjourney more, and maybe specify the resolution such as taken from a height of >75ft, as specified in the competition rules. Also there's an example image of the mannequin from the rulebook maybe you want to mimic more. We also want both lying down + standing + sitting variations, ideally:

Image

Looks good and definitely play around more!

pjoshi23 commented 11 months ago

found a dataset that has pictures of people from a bunch of different heights (5 - 30m) https://www.kaggle.com/datasets/nunenuh/semantic-drone/data