OAD-OSNet: Occlusion-Aware Dual-Branch OSNet for Occluded Person Reidentification

Despite strong performance on full-body images, person reidentification (ReID) degrades markedly when the subject is partially occluded. Monolithic architectures adapt poorly to variable occlusion, and features become corrupted by uninformative cues. We present occlusion-aware dual-branch OSNet (OAD-OSNet), which addresses occlusion through complementary dual-branch learning rather than relying solely on augmentation-driven shortcuts. OAD-OSNet retains OSNet’s efficient omniscale learning while pairing an enhanced omniscale branch with spatial–channel occlusion attention (SCOA) to suppress occlusion-consistent noise, and a complementary feature branch (CF-Branch) with global context compensation (GCC) to produce features robust to partial visibility. Channelwise branch fusion (CBF) recalibrates cross-branch channels so the two paths reinforce each other without conflating their semantics. Training uses a multitask objective with a progressive curriculum. Experiments report a strong mean average precision (mAP) on Market-1501 (90.4%) and 73.5% on MSMT17 (73.5%), competitive mAP on DukeMTMC-reID (84.2%), and the highest mAP among the methods compared on Occluded-DukeMTMC (66.5%). In comparison, Rank-1 on Occluded-DukeMTMC remains below MSCC (73.2% versus 75.0%), indicating a tradeoff between overall ranking quality and top-1 retrieval under severe occlusion.