其凭借其强大的zero-shot能力，为很多数据不足的领域也提供了帮助，突破了原有的瓶颈。
(CLIP zero-shot models) CLIP fine-tuned end-to-end CLIP fine-tuned with a linear classifier (prior work) Weight-space ensemble (end-to-end) Weight-space ensemble (linear classifier) Weight-space ensemble with α = 0.5 Standard ImageNet models Linear fit (standard ImageNet models) y=x ...
Recently, BLIP- Diffusion [28] leverages BLIP-2 [29] to align images and text for zero-shot customization. Fastcomposer [52] binds the image representation with certain text embeddings to do multiple-person generation. Some concurrent works [30, 58, 61] also explore using one referen...
动机：现有的VLM能够在自然语言提示的前提下使用zero shot open vocabulary的推理替换一组固定的支持类。...
We develop a novel method for zero shot learning (ZSL) based on test-time adaptation of similarity functions learned using training data. Existing methods exclusively employ source-domain side information for recognizing unseen classes during test time.
Plug-and-Play Sample-Efficient Fine-Tuning of Text-to-Image Diffusion Models to Learn Any Unseen StyleCVPR 2023[Paper] Target-Aware Generative Augmentations for Single-Shot AdaptationICML 2023[Paper] [Official Code] MultiDiffusion:Fusing Diffusion Paths for Controlled Image GenerationICML 2023[Paper...
Zero-Shot Semantic Segmentation Different from open-vocabulary segmentation (cross-dataset), zero-shot methods split each dataset to seen classes and unseen classes. Referring Image Segmentation Fully-Supervised Referring Image Segmentation Weakly-Supervised Referring Image Segmentation ...
Self-supervised Domain-aware Generative Network for Generalized Zero-shot Learning Jiamin Wu1, Tianzhu Zhang1,∗, Zheng-Jun Zha1, Jiebo Luo2, Yongdong Zhang1, Feng Wu1 1 University of Science and Technology of China 2 University of Rochester jiaminwu@...
Unlike these works which all use vi- sual features from classification models trained on a closed- set of pre-defined objects, places or actions, we use features from CLIP [35], which is trained on image and text paired data, and known to have strong g...
ImageNet-S 上具有不同 alpha map level的Zero-shot classification。**当foreground mask不可用时，...