1 introduction

• iccv23
• task: source-free domain generalization

common practice of DG:

1. Utilize multiple source domains for learning domain-invariant features, but it is not clear which source domains are ideal for DG.
2. Furthermore, collecting such large-scale multi-source domain data is costly.

another point:
a large-scale pre-trained model might have already seen a great variety of domains.

2 this paper

main idea: improve the model's generalization capability by simulating various distribution shifts in the latent space of such a large-scale model without using any source domain data？

Some insights:

1. as shown in Fig. motivation (a), text features could effectively represent their relevant image features.
2. the cross-model transferability phenomenon [67]: a classifier trained using text features can run inference using image features.

this paper:

• propose a prompt-driven style generation method (PromptStyler)
• synthesize styles via learnable word vectors in the form of “a S∗ style of a” (as in Fig. motivation (b))).

motivation:

imotivation

method

• the style distance between "a S* style of a" are large;
• the style-content distance between "a S* style of a [class]" are small.

image.png

Training detail:
use the learned style word vector to synthesize style-content features (synthesized features simulate images of known contents with diverse unknown styles in the joint space)

• use the text encoder to generate such style-content features
• use learnable features as input to train a linear classifier by classification loss;
• inference: image --> image encoder --> linear classifier

model：

• CLIP

exp: