Contamination MapVision / Model
MetaUAS
- Organisation
- Tencent YouTu Lab (Bin-Bin Gao)
- Family
- MetaUAS
- Released
- 2025-05-14 (arXiv v1 submission date (https://arxiv.org/abs/2505.09265); the paper appeared at NeurIPS 2024 earlier)
- Checkpoints
- csgaobb/MetaUAS metauas-256.ckpt, csgaobb/MetaUAS metauas-512.ckpt
- Weights
- Open · cc-by-nc-nd-4.0
- Links
- PaperCode
Training data disclosure: Full list
Level determined for 0% of benchmark cells.
- TestTest data seen0
- SeenSeen, split unclear0
- TrainTraining data seen0
- Related data seen0
- NoNot seen0
- ?Unknown8
“We train a universal change segmentation model on a synthetic dataset.”
“Specifically, the change segmentation dataset is synthesized using the randomly selected 60,000 images from the MS-COCO training set.”
Other declared training data
-
MS-COCO training set (60,000 images; synthetic change pairs by inpainting and Perlin-noise local changes) · split: train · fine-tuning
Not an industrial AD dataset; split refers to COCO's own train split.
Evidence
“Specifically, the change segmentation dataset is synthesized using the randomly selected 60,000 images from the MS-COCO training set.”
arxiv.org · paper · Appendix A · retrieved 2026-10-15 -
Pre-training data of base model imagenet-efficientnet-b4 (not tracked here) (broad web data: may contain benchmark test sets) · split: unknown
Evidence
“the standard convolution-based EfficientNet-b4 [57] as our encoder following UniAD [70].”
arxiv.org · paper · Section 3.2 · retrieved 2026-10-15
Benchmark cells
Industrial anomaly detection
Unknown MVTec AD · training data not disclosed in enough detail
Unknown MVTec AD 2 · training data not disclosed in enough detail
Unknown MVTec LOCO AD · training data not disclosed in enough detail
Unknown VisA · training data not disclosed in enough detail
Unknown Real-IAD · training data not disclosed in enough detail
Unknown MPDD · training data not disclosed in enough detail
Unknown BTAD · training data not disclosed in enough detail
Unknown KolektorSDD2 · training data not disclosed in enough detail
Notes
Pure vision one-prompt model: trained only on synthetic change-segmentation pairs from COCO, then evaluated on MVTec, VisA and Goods without fine-tuning; one normal prompt image per class sampled from the target normal training set. Training code not released. HF model card lists base_model google/efficientnet-b4.
Sources
- MetaUAS: Universal Anomaly Segmentation with One-Prompt Meta-Learning (NeurIPS 2024) · paper · retrieved 2026-10-15
- MetaUAS README · readme · retrieved 2026-10-15