BFS-CNN-ECA-GMP-GRU-MSP: An Enhanced Cross-Perspective GaitRecognition Model with Efficient Channel Attention and Cosine-ConsistentMetric Learning

Abstract

Cross-view gait recognition remains vulnerable to viewpoint shifts and appearance changes, especiallyunder carrying and clothing covariates. We propose BFS-CNN-ECA-GMP-GRU-MSP, an enhanced versionof our previous BFS-CNN-GMP-GRU-MSP framework, by introducing two upgrades: multi-stagelightweight channel recalibration with Efficient Channel Attention (ECA) and cosine-consistent metriclearning through cosine batch-hard triplet loss, a cosine classifier, and L2-normalized embeddings. Experimentsare first conducted on CASIA-B under the same legacy closed-set protocol used in our earlierInformatica study (gallery: NM#01–04; probe: NM#05–06, BG#01–02, CL#01–02; same-view matchesexcluded). This protocol is retained to isolate architecture-level improvements and is interpreted as awithin-protocol comparison rather than a subject-disjoint generalization benchmark. Under this setting,the proposed model reaches mean Rank-1 accuracies of 99.96% (NM), 99.74% (BG), and 98.38% (CL), improvingthe baseline by +2.96, +5.74, and +7.38 percentage points, respectively. To probe unseen-subjectbehavior more directly, we further report a supplementary subject-disjoint split (subjects 001–074 for trainingand 075–124 for testing), where the full model attains 97.73% (NM), 87.98% (BG), and 64.00% (CL).Under this stricter split, the clearest effect of ECA appears under clothing variation, where the full modelexceeds w/o ECA by 1.82 percentage points on CL, while the ECA branch still introduces only 13 learnableparameters (k=3/5/5 for 64/128/256 channels). These results support the proposed modifications asa lightweight and effective enhancement for protocol-matched cross-view gait recognition, while broadermulti-split subject-disjoint and open-set validation remains future work.