RSICRC

RSICRC: Multimodal Remote Sensing Image Change Retrieval and Captioning

RSICRC is a multimodal foundation model designed for bi-temporal remote sensing images. It jointly performs change captioning (describing changes between two images) and text-image retrieval (finding image pairs that match a text description).

The model leverages Contrastive Learning and a decoupled decoder architecture to handle both tasks simultaneously.

📄 Paper

Towards a Multimodal Framework for Remote Sensing Image Change Retrieval and Captioning Roger Ferrod, Luigi Di Caro, Dino Ienco Published at Discovery Science 2024

Read the Paper | GitHub Repository

🏗️ Model Architecture

The framework is inspired by CoCa but adapted for bi-temporal remote sensing data.

• Encoder: A Siamese network (ResNet-50 or ViT via OpenCLIP) that encodes "before" and "after" images. A Hierarchical Self-Attention (HSA) block and a residual block with a cosine mask fuse the bi-temporal features.
• Decoder: A decoupled Transformer decoder split into:
  • Unimodal Layers: Encode text only (used for contrastive alignment).
  • Multimodal Layers: Apply cross-attention between visual and textual features to generate captions.

💻 Usage

To use this model use the custom source code.

Inference Code

import torch
import json
import open_clip
from huggingface_hub import hf_hub_download

from src.model import ICCModel 

# 1. Download necessary files from Hugging Face
repo_id = "rogerferrod/RSICRC"
config_path = hf_hub_download(repo_id=repo_id, filename="config.json")
vocab_path = hf_hub_download(repo_id=repo_id, filename="levir_vocab.json") 
weights_path = hf_hub_download(repo_id=repo_id, filename="pytorch_model.bin")

# 2. Load Configuration and Vocabulary
with open(config_path, 'r') as f:
    config = json.load(f)

with open(vocab_path, 'r') as f:
    vocab = json.load(f)

# 3. Setup Device and Backbone (OpenCLIP)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
clip_model, _, preprocess = open_clip.create_model_and_transforms(config['backbone'])

# 4. Initialize the Model
model = ICCModel(
    device=device,
    clip=clip_model,
    backbone=config['backbone'],
    d_model=config['d_model'],
    vocab_size=len(vocab),
    max_len=config['max_len'],
    num_heads=config['num_heads'],
    h_dim=config['h_dim'],
    a_dim=config['a_dim'],
    encoder_layers=config['encoder_layers'],
    decoder_layers=config['decoder_layers'],
    dropout=config['dropout'],
    learnable=config['learnable'],
    fine_tune=config['fine_tune'],
    tie_embeddings=config['tie_embeddings'],
    prenorm=config['prenorm']
)

# 5. Load Weights
model.load_state_dict(torch.load(weights_path, map_location=device))
model = model.to(device)
model.eval()

print("Model loaded successfully!")

📚 Citation

If you use this model or code in your research, please cite our paper:

@InProceedings{10.1007/978-3-031-78980-9_15,
author       = {Roger Ferrod and
                  Luigi Di Caro and
                  Dino Ienco},
editor       = {Dino Pedreschi and
                Anna Monreale and
                Riccardo Guidotti and
                Roberto Pellungrini and
                Francesca Naretto},
title        = {Towards a Multimodal Framework for Remote Sensing Image Change Retrieval
                and Captioning},
booktitle    = {Discovery Science - 27th International Conference, {DS} 2024, Pisa,
                Italy, October 14-16, 2024, Proceedings, Part {II}},
series       = {Lecture Notes in Computer Science},
volume       = {15244},
pages        = {231--245},
publisher    = {Springer},
year         = {2024},
url          = {https://doi.org/10.1007/978-3-031-78980-9\_15},
doi          = {10.1007/978-3-031-78980-9\_15}
}