Causal Bootstrapped Alignment for Unsupervised Video-Based Visible-Infrared Person Re-Identification
A new method uses causal AI to match people across visible and infrared video without expensive labeled data.
A research team has introduced a novel AI framework called Causal Bootstrapped Alignment (CBA) that solves a major hurdle in 24/7 surveillance: matching a person's identity across visible-light and infrared video feeds without any manually labeled training data. This task, known as Unsupervised Video-Based Visible-Infrared Person Re-Identification (USL-VVI-ReID), is critical for all-day security but has been limited by the high cost of cross-modality annotations. Existing methods that simply adapt image-based techniques to video perform poorly, as generic AI encoders get confused by irrelevant details like clothing motion and fail to align the different 'granularity' of data from the two camera types.
The CBA framework directly attacks these problems with a two-stage, causality-inspired approach. First, its Causal Intervention Warm-up (CIW) module treats video sequences as a source of inherent prior knowledge. It performs sequence-level interventions to suppress spurious correlations caused by modality differences and movement, isolating the core, identity-relevant information for more reliable AI clustering. Second, the Prototype-Guided Uncertainty Refinement (PGUR) module tackles the alignment mismatch. It uses reliably clustered identities from the visible spectrum as guides to reorganize the more challenging infrared data, applying uncertainty-aware supervision for a fine-grained match.
Extensive testing on standard benchmarks like HITSZ-VCM and BUPTCampus shows CBA significantly outperforms previous unsupervised methods adapted to the video setting. This represents a major step toward practical, large-scale deployment of cross-modality tracking systems that can learn directly from unlabeled operational data, bypassing the annotation bottleneck that has constrained their scalability.
- Eliminates need for costly labeled data by using an unsupervised learning approach for cross-modality person tracking.
- Uses causal intervention techniques to filter out irrelevant motion and modality noise, improving clustering accuracy by 15-20% on benchmarks.
- Enables scalable 24/7 surveillance systems that can automatically correlate identities between daytime (visible) and nighttime (infrared) footage.
Why It Matters
This breakthrough reduces the cost and complexity of deploying AI-powered, round-the-clock security and monitoring systems at scale.