Open-Vocabulary Camouflaged Object Segmentation with Cascaded Vision-Language Models

---

📌 Overview

This repository provides the official implementation of our paper:
"Open-Vocabulary Camouflaged Object Segmentation with Cascaded Vision-Language Models",
which introduces a cascaded two-stage framework for segmenting and recognizing camouflaged objects in open-vocabulary settings.

---

📁 Setup

🔹 Dataset

1. Download the OVCamo dataset from the official repository.
2. Update the dataset path in the following config file:

./datasets/ovcamo_info/splitted_ovcamo.yaml

🔹 Pretrained Model

Download the SAM backbone from Meta AI:

• sam_vit_h_4b8939.pth

Place the file in the ./pretrained directory:

./pretrained/sam_vit_h_4b8939.pth

---

🚀 Demo

Download our best pre-trained model:

• model_epoch_best.pth
• model-best.pth.tar

Save it to:

./best_model_pth/model_epoch_best.pth

Run the demo with:

python demo.py \
  --img-path ./demo_img/scorpionfish.jpg \
  --output-dir ./demo_img \
  --config ./configs/demo.yaml \
  --model ./best_model_pth/model_epoch_best.pth