<p>This paper focuses on the challenge of learning compact and distinctive image embeddings for single-stage large-scale visual place recognition (VPR). While conventional two-stage methods suffer from high computational overhead due to local feature re-ranking, and single-stage solutions often depend on complex multi-branch architectures for local–global fusion, we introduce AdaArc (Adaptive Weighted Bidirectional Angular Margin Loss)–a novel loss function that streamlines feature optimization. AdaArc integrates two synergistic mechanisms: a) bidirectional angular margin constraints, which impose simultaneous angular-space margins on both target and non-target classes to enhance intra-class compactness while suppressing inter-class similarity, and b) adaptive weighting, which dynamically prioritizes hard and easy samples based on their proximity to class centroids to optimize training focus. Powered solely by AdaArc and a standard ResNet50 backbone, our framework achieves state-of-the-art discriminability among single-stage, ResNet50-based architectures in large-scale VPR. Experiments demonstrate the superiority of our solution over existing methods across key benchmarks, while reducing temporal and spatial costs. This confirms that complex architectures are unnecessary for high-performance VPR when theoretically grounded losses like AdaArc orchestrate feature space optimization. Although designed for streamlined models, auxiliary experiments validate AdaArc’s compatibility with mainstream single-stage models.</p>