Children acquire object category representations from their everyday experiences in the first few years of life.
What do the inputs to this learning process actually look like?
New preprint! arxiv.org/abs/2605.14990
What did we find? First, we found that a few categories dominate — across all domains! Children see some categories much more often than others, replicating prior work.
Content: huge variation in viewpoint & format — including lots of depictions (esp. animals—examples below!!.
Children are learning about many categories from these kinds of “child-directed” visuals rather than real-life exemplars.
Category organization: We found stronger within-superordinate clustering in the child’s view than in a curated dataset (THINGS). We also find this pattern in data from individual children.
Maybe this provides a way for children to parse the “blooming buzzing confusion” of their visual world!
I’m giving a talk on this at #VSS2026: Sun May 17, 8:30am, Talk Room 2 (Object Recognition, 31.22). My first VSS! www.visionsciences.org/talk-session...
Thanks to coauthors @tarunsepuri.bsky.social @alvinwmtan.bsky.social, @khaiaw.bsky.social, @mcxfrank.bsky.social, and @brialong.bsky.social!
We analyzed the objects in view in 868 hours of egocentric videos from the BabyView dataset—from 31 children (5-36 months old). We used an object detection model to extract object categories from >3M frames—pipeline below.
And more about the BabyView Project here! babyview-project.github.io
We are still hiring for our computational focus lab coordinator position!
If you have a computational / software engineering background and are looking for more research experience before applying to graduate school, this would be a great fit.
Read more about our lab here!
www.vislearnlab.org