Luggage Re-Identification via Mask-Guided Spatial Attention Network

Accurately matching luggage across different camera views is essential for both airport security operations and passenger services, but this task poses substantial difficulties owing to the striking resemblance among various baggage items and the complex, cluttered scenes typical of airport settings. We propose a new framework aimed at improving feature distinctiveness for Luggage ReID. The proposed framework employs a ResNet101 network as the backbone feature extractor, trained with a joint objective comprising triplet loss - which refines the embedding space by minimizing distances between same-identity pairs while maximizing separation from different identities - and cross-entropy classification loss. Central to our contribution is a novel Mask-Guided Spatial Attention Network (MSAN), which steers the network's attention toward the core luggage region, where appearance characteristics tend to be most consistent and distinctive. Specifically, we construct a smooth Gaussian-weighted attention template centered on the image as a spatial guidance signal. A dedicated decoder branch within MSAN transforms backbone features into a predicted attention map, supervised via Mean Squared Error loss with respect to the Gaussian template. This mechanism encourages the network to emphasize discriminative luggage structures while attenuating distracting background elements. Comprehensive evaluations on standard benchmarks confirm that our approach yields more resilient and distinguishable representations, attaining leading performance in luggage re-identification.