bi-modal textual prompt learning for vision-language models in remote sensing

TL;DR

BiMoRS tackles domain generalization in remote sensing by creating image-adaptive prompts that fuse textual descriptions from a frozen image captioning model with visual cues from a frozen VLM (CLIP). A lightweight cross-attention module conditions learnable prompt tokens on this bi-modal context, producing contextualized prompts without altering the CLIP/BLIP-2 backbones. The approach yields consistent gains across base-to-new, cross-dataset, and single-source multi-target settings, with up to a 2% improvement and substantially fewer trainable parameters. This work demonstrates the effectiveness of bi-modal, image-conditioned prompting for RS tasks and points to future extensions to segmentation and retrieval.

Abstract

Prompt learning (PL) has emerged as an effective strategy to adapt vision-language models (VLMs), such as CLIP, for downstream tasks under limited supervision. While PL has demonstrated strong generalization on natural image datasets, its transferability to remote sensing (RS) imagery remains underexplored. RS data present unique challenges, including multi-label scenes, high intra-class variability, and diverse spatial resolutions, that hinder the direct applicability of existing PL methods. In particular, current prompt-based approaches often struggle to identify dominant semantic cues and fail to generalize to novel classes in RS scenarios. To address these challenges, we propose BiMoRS, a lightweight bi-modal prompt learning framework tailored for RS tasks. BiMoRS employs a frozen image captioning model (e.g., BLIP-2) to extract textual semantic summaries from RS images. These captions are tokenized using a BERT tokenizer and fused with high-level visual features from the CLIP encoder. A lightweight cross-attention module then conditions a learnable query prompt on the fused textual-visual representation, yielding contextualized prompts without altering the CLIP backbone. We evaluate BiMoRS on four RS datasets across three domain generalization (DG) tasks and observe consistent performance gains, outperforming strong baselines by up to 2% on average. Codes are available at https://github.com/ipankhi/BiMoRS.

Figures (2)

• Figure 1: Model architecture of BiMoRS, comprises of two main components: (i) firstly, we propose to capture the visual information through an image-captioning model $\mathcal{G}_v$ as attribute descriptions and tokenize them using a tokenizer $\mathcal{G}_t$, and (ii) secondly, we propose a lightweight cross-attention module $\bm{\mathrm{CA}}$, that effectively combines both of the visual features from the final layer of $\mathcal{F}_v$ and tokenized descriptions, rescaled through projectors $P_v$ and $P_t$ respectively. Finally, $\bm{\mathrm{CA}}$ generates bi-modal text prompts that efficaciously generalize BiMoRS through textual prompting with the better alignment of V-L representations.
• Figure 2: Comparison of attention maps generated by SOTA methods on a sample of class 'Harbor' using Grad-CAM gradcam.