Deep Multimodal Fusion for Generalizable Person Re-identification