Recall

website page

Inspiration

Reminiscence therapy is standard in dementia care, but the medium hasn't changed in decades. A photo is a flat window into a life you can't step back into. We wanted to see what happens when a memory becomes somewhere you can stand.

What it does

Recall turns ordinary photos and video of a meaningful place into a navigable 3D scene using Gaussian splatting. A caregiver uploads whatever footage the family has, and the user walks through the reconstructed space at their own pace. Not photorealism for its own sake, just a familiar room that feels like theirs.

How we built it

Gaussian splatting for reconstruction, Unity for rendering, navigation, and UI. Animation work went into making movement gentle and legible instead of the default free fly, which matters for this audience. AI models handled segmentation on sparse, low quality source footage.

Challenges we ran into

Open source vision models aren't ready for real input. They hold up on curated benchmark examples and fall apart on motion blur, bad lighting, and the partial captures family archives actually consist of. We also couldn't lean on generative fill: for a dementia user, a hallucinated detail isn't an artifact, it's a false memory. We reconstructed conservatively and left gaps visible.

Accomplishments that we're proud of

A full working app, capture to reconstruction to walkthrough, built in the hackathon window. Nothing in the demo is stubbed.

What we learned

Coding is not solved. Scaffolding came fast, but integration, silent reconstruction failures, and performance on consumer hardware ate the time. We also learned that building for a vulnerable user changes the technical calculus. Accuracy stopped being a quality metric and became an ethical one.

What's next for Recall

Web deployment, so a family can upload footage without installing anything or owning a GPU. After that, phone based capture and testing with memory care clinicians to see whether this actually beats photographs for reminiscence.