This repository provides SINet, a strong baseline for Camouflaged Object Detection (COD), a challenging computer vision task focused on identifying objects that blend seamlessly with their environment. It's aimed at researchers and practitioners in computer vision, offering a robust model, a comprehensive dataset (COD10K), and evaluation tools for advancing COD research.

How It Works

SINet employs a novel architecture featuring a receptive field (RF) component to mimic human visual processing and a partial decoder component (PDC) that simulates animal predation's search and identification stages. This dual approach allows the model to effectively capture subtle visual cues and contextual information necessary for detecting camouflaged objects, outperforming traditional salient object detection methods.

Quick Start & Requirements

• Install: Create a conda environment (conda create -n SINet python=3.6), activate it, and install dependencies (pip install -r requirements.txt).
• Prerequisites: Ubuntu OS, Python 3.6, PyTorch. NVIDIA Apex (optional, for accelerated training with mixed precision) requires CUDA 10.0 and Cudnn 7.4.
• Data: Download COD10K training and testing datasets from provided links.
• Testing: Run MyTest.py with specified model and save paths.
• Evaluation: Use MATLAB scripts in EvaluationTool for performance metrics.
• Links: Project Page, Manuscript (PDF), Chinese Version (PDF), Online Demo.

Highlighted Details

• SINet is the best-performing method on the open benchmark website (paperswithcode.com).
• An enhanced version, SINet-V2, achieves real-time inference and surpasses existing COD methods.
• The COD10K dataset offers detailed annotations, including edge-level, object-level, and instance-level data.
• A MATLAB-based evaluation toolbox is provided for standardized benchmarking.

Maintenance & Community

The project is associated with CVPR 2020 (Oral) and has a WeChat group for community discussion. A newer version (SINet-V2) is accepted at IEEE TPAMI 2022. Training code availability is via email request for research purposes.

Licensing & Compatibility

The code is for academic communication and research purposes only. The COD10K Dataset is strictly for non-commercial use. Commercial use requires direct contact with the authors.

Limitations & Caveats

The code is primarily tested on Ubuntu OS and may not be guaranteed to work on other platforms. The README indicates plans to support various backbones, distributed training, and lightweight architectures, suggesting these are not yet implemented.