Kaleidoscopic Background Attack: Disrupting Pose Estimation
with Multi-Fold Radial Symmetry Textures

Xinlong Ding · Hongwei Yu · Jiawei Li · Feifan Li · Yu Shang · Bochao Zou · Huimin Ma · Jiansheng Chen

University of Science and Technology Beijing   |   Tsinghua University

🏆 Overview

Official PyTorch implementation for ICCV 2025 Paper: Kaleidoscopic Background Attack: Disrupting Pose Estimation with Multi-Fold Radial Symmetry Textures.

🚀 Installation

To get started, please make sure your environment meets the following requirements:

• GPU with at least 24GB memory (we recommend ~33GB)
• Ubuntu 22.04, CUDA 12.4
• Python >= 3.10

Step 1: Clone and Set Up Python Environment

git clone --recursive https://github.com/wakuwu/KBA
cd KBA

# Install uv (https://docs.astral.sh/uv/)
uv sync

# Install PyTorch3D (CUDA 12.4 compatible)
uv pip install --extra-index-url https://miropsota.github.io/torch_packages_builder pytorch3d==0.7.8+pt2.4.1cu124

Step 2: Color Management Setup

Install required tools:

sudo apt-get update
sudo apt-get install liblcms2-dev liblcms2-utils

Then download the Adobe ICC Profiles, accept the license, and unzip the archive AdobeICCProfilesCS4Win_bundler. Copy the CMYK folder into the following directory:

data/cms/

Step 3: Prepare Data

Download our preprocessed attack dataset:

wget https://huggingface.co/datasets/umiskky/KBA/resolve/main/data.tar
tar -xf data.tar

You can also optionally download:

• OmniObject3D models from OpenXLab, placed under data/dataset/
• HDRI environment maps from PolyHaven, placed under data/environments/

Step 4: DUSt3R Configuration

Download the DUSt3R pretrained weights:

DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth

Place the downloaded file in the checkpoint directory:

third_party/dust3r/checkpoints/

🐳 Optional: Run with Docker

We also provide a pre-built Docker image for convenience:

docker pull ghcr.io/wakuwu/kba:latest

🔬 Evaluation & Rendering

After setup, you can test the system using the following commands:

# Run DUSt3R pose estimation and 3D reconstruction
python third_party/dust3r/demo.py \
    --weights third_party/dust3r/checkpoints/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth

# Render multi-view images with specified kaleidoscopic background
python test.py

🛡️ Launch Attack

To launch the kaleidoscopic background attack:

python attack_dust3r.py

📜 Citation

If you find this work helpful, please consider citing our paper:

@article{ding2025kba,
    title   = {Kaleidoscopic Background Attack: Disrupting Pose Estimation with Multi-Fold Radial Symmetry Textures},
    author  = {Xinlong Ding, Hongwei Yu, Jiawei Li, Feifan Li, Yu Shang, Bochao Zou, Huimin Ma and Jiansheng Chen},
    journal = {arXiv preprint arXiv:2507.10265},
    year    = {2025}
}

📄 License

This project is licensed under the Apache License 2.0. See the LICENSE file for more details.

🌠 Star History