Taming Generative Synthetic Data for X-ray Prohibited Item Detection

Quick Overview

The X-ray synthesis technique successfully tames generative data by intelligently mixing real (low-quality/high-cost) and synthetic (high-quality/low-cost) data, achieving a 70.7% AP improvement over previous manual methods while avoiding the pitfalls of over-specialization.

Key Points: The X-ray synthesis method achieves a 70.7% AP improvement over previous labor-intensive methods for detecting prohibited items like knives and guns in security scans. The key innovation involves a two-stage pipeline: first, generating high-quality synthetic data based on real X-ray images, and second, fusing this synthetic data with the original real data to create a refined training set. Previous methods, whether tip-based or GAN-based, were stuck in a two-stage pipeline requiring extensive manual annotation or resulting in models that overfit to the training data. The X-ray synthesis approach uses a cross-attention map to precisely determine which pixels from the synthetic image should be fused with the real image, avoiding unrealistic artifacts. This fusion technique successfully filters out noise and prevents the model from overfitting to the clean synthetic data, which would fail in real-world, cluttered scenarios. The process avoids the high cost and labor of manually annotating every prohibited item in every image, which was a major bottleneck in prior security inspection training. The resulting model is highly generalizable, performing well on real X-ray data (like CT scans) even when trained primarily on synthetic data, demonstrating robustness against data scarcity.

Context: This video discusses a novel approach to training AI models for detecting prohibited items (like weapons) in security X-ray or CT scans, addressing the significant challenge of data scarcity and the high cost of manual annotation for this critical safety application. The presenters contrast older, less effective methods with their new, two-stage synthesis technique designed to create high-quality, realistic training data that generalizes well to real-world scenarios.