Skip to main content
AIMSAI Measurement Science

General

Visual Memory (flexible perception)

Per-test-image top-1 nearest-neighbor retrieval correctness for 5 embedding models (DinoV2 ViT-S/B/L14, CLIP ViT-B16/L14) using an ImageNet-2012 visual memory, on three ImageNet distribution-shift test sets (ImageNet-A/R/V2). Released as per-image nearest-neighbor JSON on gs://visual_memory_v01.

47,500items
5subjects
100%observed
cc-by-4.0license
generaldomain
imagemodality
Original sourcePaperBuild script← All benchmarks

Response matrix

Fit to width. Hover for subject & item; click a cell for details.

Visual Memory (flexible perception) response matrix: AI models (rows) against items (columns)
Correct (1)Incorrect (0)Unobserved

Scale: 1 = correct · 0 = incorrect

Subjects

  1. 1DinoV2 ViT-L/140.5743
  2. 2DinoV2 ViT-B/140.4839
  3. 3CLIP ViT-L/140.4729
  4. 4DinoV2 ViT-S/140.3703
  5. 5CLIP ViT-B/160.334
Full data on Hugging FaceBack to the gallery