CORE - Persistent, Compositional, Object Memory for Vision & Embodied AI

Your frozen vision model (CLIP, Llava, SigLIP) sees the world, but forgets it instantly. CORE gives it memory that lasts.
✓ Persistent Object Identity - Assigns unique, updatable IDs to every object, so your robot knows it's the same cup, even after 100 steps or a reboot.
✓ Compositional Understanding - Remembers that a chair has legs and a seat, and how they connect. Enables zero-shot assembly and repair.
✓ Lightweight - No retraining. No regression. Just +1 matrix multiply on your frozen backbone.

Connecting to server...

Kernels: -- Sessions: --

Object Detection & Recognition

Upload an image to detect and identify objects

📸

Drop image here or click to upload

Supports JPG, PNG, GIF

Text Description (optional - helps with detection)

Session ID (optional - for tracking)

Analyzing image...

Find similar images from a collection

Reference Image

🖼️

Upload reference image

Candidate Images (upload multiple)

🗂️

Upload images to compare

Top K Results

Computing similarities...

Track objects across multiple frames

Tracking Session ID

🎥

Upload frame image

Upload multiple images sequentially for the same session

Tracking objects...

Understand how the model works internally

Analysis Type

Analyzing model internals...

Explore object relationships and hierarchies

Analyzing hierarchical structure...

Test the complete pipeline with one click

Running complete pipeline test...