Swapping Faces, Saving Features: A Dual-Purpose Pipeline for Pedestrian Privacy in ITS

2026-07-11 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Cybersecurity & Data Privacy · Depth: Expert, long

Summary

A five-stage pipeline has been developed to protect pedestrian privacy in Intelligent Transportation Systems (ITS) datasets while maintaining essential facial attributes for AI model training. Tailored for urban Egyptian images, the pipeline integrates Pedestrian Detection (YOLOV11), Face Detection (SCRFD), Quality Enhancement (Codeformer), Face Swapping, and Blending (OpenCV's SeamlessClone). The work evaluated Roop and Ghost-v2 face-swapping models, finding Roop superior in preserving facial expressions, structure, and pose, and more robust in challenging scenarios like occluded faces or veiled women. Ghost-v2 showed better identity concealment but struggled with realism and attribute preservation in complex cases. The pipeline successfully maintained critical features like "looking/not looking" for downstream tasks such as pedestrian intention prediction.

Key takeaway

For Machine Learning Engineers developing privacy-preserving datasets for autonomous vehicles, consider implementing a multi-stage face-swapping pipeline. This approach effectively conceals pedestrian identities while preserving crucial facial attributes like gaze and expression for model training. You should prioritize models like Roop, which demonstrated superior robustness and attribute preservation over Ghost-v2, especially with diverse urban imagery and low-resolution inputs. Integrate quality enhancement for optimal results.

Key insights

Face swapping offers a balanced approach to pedestrian privacy and data usability for ITS AI model training.

Principles

• Privacy methods must preserve data usability.
• Face swapping maintains natural facial structure.
• Robustness is key for real-world urban datasets.

Method

A five-stage pipeline: pedestrian detection (YOLOV11), face detection (SCRFD), quality enhancement (Codeformer), face swapping (Roop), and blending (OpenCV SeamlessClone).

In practice

• Use Roop for robust face swapping.
• Integrate Codeformer for low-res image quality.
• Evaluate identity concealment via cosine similarity.

Topics

• Intelligent Transportation Systems
• Pedestrian Privacy
• Face Swapping
• Autonomous Vehicles
• Computer Vision
• Data Anonymization

Code references

• ultralytics/ultralytics
• s0md3v/roop
• TencentARC/GFPGAN

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Swapping Faces, Saving Features: A Dual-Purpose Pipeline for Pedestrian Privacy in ITS — A five-stage face-swapping pipeline balances pedestrian privacy and data usability for autonomous vehicle training datasets, with Roop outperforming Ghost-v2.
• Adaptive Cross-Modal Fusion with Sparse Attention for Pedestrian Crossing Intention Prediction — ADAPT uses sparse cross-modal attention and multimodal fusion for efficient, accurate pedestrian intention prediction.
• Realistic Face Reconstruction from Facial Embeddings via Diffusion Models — FEM reconstructs high-resolution faces from embeddings, exposing privacy risks in FR and PPFR systems.
• Benchmarking Face Recognition without Real Faces — Well-constructed synthetic datasets can reliably evaluate face recognition models, enabling fully privacy-preserving pipelines.
• CityGuard: Graph-Aware Private Descriptors for Bias-Resilient Identity Search Across Urban Cameras — CityGuard enables bias-resilient, privacy-preserving person re-identification by integrating geometry-aware graph attention, adaptive metric learning, and differential privacy.
• Towards High Fidelity Face Swapping: A Comprehensive Survey and New Benchmark — Standardized benchmarks and structured surveys are crucial for advancing face swapping research.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.