---
pipeline_tag: object-detection
---
# MSRNet: A Multi-Scale Recursive Network for Camouflaged Object Detection.
This repository is for MSRNet: A Multi-Scale Recursive Network for Camouflaged Object Detection, introduced in this [paper](https://huggingface.co/papers/2511.12810).

The model is built using PyTorch on an NVIDIA RTX A6000 GPU with a total memory size of 48 GB.
The code is also available in our [GitHub](https://github.com/linaagh98/MSRNet) repository.

## Contents
1. [Introduction](#introduction)
2. [Network](#network)
4. [Data Preparation](#data-preparation)
5. [Requirements Installation](#requirements-installation)
6. [Training](#training)
7. [Evaluation](#evaluation)
8. [Results](#results)
9. [Acknowledgement](#acknowledgment)

## Introduction
Camouflaged object detection (COD) is a challenging computer vision task that aims to identify and segment objects blending into their environment. MSRNet proposes a Multi-Scale Recursive Network to extract and combine multi-scale features, using a Pyramid Vision Transformer backbone and Attention-Based Scale Integration Units. Its decoder refines features recursively with Multi-Granularity Fusion Units and a novel recursive-feedback decoding strategy to enhance global context. This approach improves detection of small and multiple camouflaged objects, achieving state-of-the-art results on several benchmark datasets.

## Network
This diagram illustrates the overall architecture of MSRNet. 
![Methodology](Images/MethodologyDiagram.png)

## Data Preparation
In this research, we utilized four benchmark datasets for camouflaged object detection (CAMO, CHAMELEON, COD10K, and NC4K).

After downloading all datasets, you need to create a file named "dataset.yaml" and place it in the same directory as the main code folder.  

The dataset.yaml file will include the paths for your Train and Test datasets. Please ensure that you place the datasets in the corresponding paths as you specified in the dataset.yaml file. 

Your dataset.yaml file should look something like this:

```yaml
# ICOD Datasets
cod10k_tr:
  {
    root: "YOUR_ROOT_DIRECTRY/ICOD_Datasets/Train/COD10K-TR",
    image: { path: "Image", suffix: ".jpg" },
    mask: { path: "Mask", suffix: ".png" },
  }
camo_tr:
  {
    root: "YOUR_ROOT_DIRECTRY/ICOD_Datasets/Train/CAMO-TR",
    image: { path: "Image", suffix: ".jpg" },
    mask: { path: "Mask", suffix: ".png" },
  }
cod10k_te:
  {
    root: "YOUR_ROOT_DIRECTRY/ICOD_Datasets/Test/COD10K-TE",
    image: { path: "Image", suffix: ".jpg" },
    mask: { path: "Mask", suffix: ".png" },
  }
camo_te:
  {
    root: "YOUR_ROOT_DIRECTRY/ICOD_Datasets/Test/CAMO-TE",
    image: { path: "Image", suffix: ".jpg" },
    mask: { path: "Mask", suffix: ".png" },
  }
chameleon:
  {
    root: "YOUR_ROOT_DIRECTRY/ICOD_Datasets/Test/CHAMELEON",
    image: { path: "Image", suffix: ".jpg" },
    mask: { path: "Mask", suffix: ".png" },
  }
nc4k:
  {
    root: "YOUR_ROOT_DIRECTRY/ICOD_Datasets/Test/NC4K",
    image: { path: "Imgs", suffix: ".jpg" },
    mask: { path: "GT", suffix: ".png" },
  }
```
## Requirements Installation

* torch==2.1.2
* torchvision==0.16.2
* Others: `pip install -r requirements.txt`

## Training

```shell
python main_for_image.py --config configs/icod_train.py --pretrained --model-name EffB1_MSRNet
python main_for_image.py --config configs/icod_train.py --pretrained --model-name EffB4_MSRNet
python main_for_image.py --config configs/icod_train.py --pretrained --model-name PvtV2B2_MSRNet
python main_for_image.py --config configs/icod_train.py --pretrained --model-name PvtV2B3_MSRNet
python main_for_image.py --config configs/icod_train.py --pretrained --model-name PvtV2B4_MSRNet
python main_for_image.py --config configs/icod_train.py --pretrained --model-name PvtV2B5_MSRNet
python main_for_image.py --config configs/icod_train.py --pretrained --model-name RN50_MSRNet
```
> [!note]
> These command-lines will not save the final predection images of the trained model, to save the predection sesults of your traind model add --save-results to your command-line.


## Evaluation

```shell
python main_for_image.py --config configs/icod_train.py --model-name <MODEL_NAME> --evaluate --load-from <TRAINED_WEIGHT>
```


## Results
The prediction results of our highest-performing model (PvtV2B4_MSRNet) on [CAMO](https://drive.google.com/drive/folders/12HoC5l_0gL_JjpCIBHO5sevMLTbZT6zP?usp=drive_link), [CHAMELEON](https://drive.google.com/drive/folders/1p0i6y3seR0a_RBPzlXYIKxaif3A9ZytT?usp=drive_link), [COD10K](https://drive.google.com/drive/folders/10yR26pNG4La7ikNJTKmu9qGLJp5xwcYS?usp=drive_link), and [NC4K](https://drive.google.com/drive/folders/19xBkQUVZ597n8Ilav3QKsBRvQVhf3WOQ?usp=drive_link) are available, along with the [model weights](https://drive.google.com/file/d/12M_Cw9B96z9QpeGq5wRPW5ZGJlavJs0L/view?usp=drive_link). 

### MSRNet Performance Results

| Backbone        | CAMO  |       |       |       |       |CHAMELEON|       |       |       |       | COD10K |       |       |       |       | NC4K  |       |       |       |       |
| --------------- | ----- | ------| ----- | ------|-------|---------| ----- | ----- | ------|-------| -------| ----- | ----- | ------|-------| ----- | ----- | ----- |-------|-------|
|                 |  S_m  | F^ω_β |  MAE  |  F_β  |  E_m  |   S_m   | F^ω_β | MAE   |  F_β  |  E_m  |  S_m   | F^ω_β | MAE   |  F_β  |  E_m  |  S_m  | F^ω_β | MAE   |  F_β  |  E_m  |
| ResNet-50       | 0.816 | 0.754 | 0.071 | 0.794 | 0.872 |  0.918  | 0.876 | 0.020 | 0.888 | 0.975 | 0.868  | 0.786 | 0.024 | 0.816 | 0.934 | 0.869 | 0.814 | 0.039 | 0.844 | 0.925 |
| EfficientNet-B4 | 0.875 | 0.838 | 0.045 | 0.863 | 0.936 |  0.923  | 0.881 | 0.019 | 0.891 | 0.970 | 0.887  | 0.814 | 0.020 | 0.838 | 0.947 | 0.889 | 0.844 | 0.031 | 0.866 | 0.943 |
| PVTv2-B2        | 0.873 | 0.838 | 0.047 | 0.860 | 0.928 |  0.931  | 0.904 | 0.016 | 0.912 | 0.976 | 0.894  | 0.829 | 0.018 | 0.849 | 0.952 | 0.894 | 0.853 | 0.030 | 0.874 | 0.943 |
| PVTv2-B3        | 0.885 | 0.855 | 0.043 | 0.874 | 0.941 |  0.933  | 0.907 | 0.016 | 0.915 | 0.973 | 0.904  | 0.847 | 0.017 | 0.865 | 0.959 | 0.903 | 0.867 | 0.027 | 0.886 | 0.952 |
| PVTv2-B4        | 0.888 | 0.861 | 0.040 | 0.878 | 0.942 |  0.932  | 0.908 | 0.017 | 0.916 | 0.978 | 0.907  | 0.852 | 0.016 | 0.868 | 0.962 | 0.905 | 0.873 | 0.026 | 0.890 | 0.953 |
| PVTv2-B5        | 0.888 | 0.860 | 0.041 | 0.876 | 0.943 |  0.925  | 0.893 | 0.017 | 0.903 | 0.971 | 0.902  | 0.844 | 0.017 | 0.862 | 0.957 | 0.903 | 0.871 | 0.027 | 0.889 | 0.952 |


### MSRNet Visual Results

A visual Comparison showing the superiority of MSRNet in detecting multiple (rows 1-3), small (rows 4 and 5), and tiny (rows 6 and 7) camouflaged objects.
![visual-Comparison](Images/VisualComparison.png)






 

## Acknowledgment
This project builds upon the work of [Lart Pang](https://github.com/lartpang), incorporating key modifications to the decoding strategy and input scales to enhance its ability to detect small and multiple camouflaged objects.