Contamination Map Methodology

Vision / Model

AA-CLIP, trained on VisA

Organisation
Zhejiang University et al.
Family
AA-CLIP
Released
2025-03-09 (arXiv v1 submission date (https://arxiv.org/abs/2503.06661))
Weights
Closed · Apache-2.0
Links
PaperCode

Training data disclosure: Full list

Level determined for 12% of benchmark cells.

“We train our model on a real-world industrial AD dataset - VisA [60] - in which objects are different from other datasets.”
arxiv.org · paper · Section 4.1 · retrieved 2026-10-15
“we conduct training under various data levels: 2-shot per class, 16-shot per class, 64-shot per class, and full-shot.”
arxiv.org · paper · Section 4.1 · retrieved 2026-10-15

Built on

Everything these base models may have seen is inherited by AA-CLIP, trained on VisA.

Other declared training data

Benchmark cells

Industrial anomaly detection

Unknown MVTec AD · unknown for its base model OpenAI CLIP: training data not disclosed in enough detail

Unknown MVTec AD 2 · unknown for its base model OpenAI CLIP: training data not disclosed in enough detail

Unknown MVTec LOCO AD · unknown for its base model OpenAI CLIP: training data not disclosed in enough detail

Seen, split unclear VisA · in fine-tuning data: the dataset was used; the source does not say which part

Evidence (2)
“We train our model on a real-world industrial AD dataset - VisA [60] - in which objects are different from other datasets.”
arxiv.org · paper · Section 4.1 · retrieved 2026-10-15
“we conduct training under various data levels: 2-shot per class, 16-shot per class, 64-shot per class, and full-shot.”
arxiv.org · paper · Section 4.1 · retrieved 2026-10-15

Unknown Real-IAD · unknown for its base model OpenAI CLIP: training data not disclosed in enough detail

Unknown MPDD · unknown for its base model OpenAI CLIP: training data not disclosed in enough detail

Unknown BTAD · unknown for its base model OpenAI CLIP: training data not disclosed in enough detail

Unknown KolektorSDD2 · unknown for its base model OpenAI CLIP: training data not disclosed in enough detail

Notes

Configuration (no released weights found; README asks users to train). Two-stage residual adapters on a frozen CLIP. VisA-trained models (2/16/64-shot and full-shot) are used for results on all datasets except VisA (MVTec AD, BTAD, MPDD and medical datasets).

Sources