Rethinking the Aggregation Head: Competitive Slot Distillation for Robust Visual Place Recognition

Visual Place Recognition (VPR) localizes a query image by retrieving the best-matching place from a large-scale geo-tagged database. Its key challenge lies in aggregating the stable structural cues from distractor-saturated scenes into a global descriptor. Existing aggregation methods follow a single-pass paradigm that struggles to fully disentangle informative cues from distractors, and largely offload this unresolved burden onto auxiliary external modules. Inspired by natural selection, we recast aggregation as an iterative structure-distillation process, in which stable structural cues are progressively reinforced through competition, while distractor responses are suppressed to guide later iterations toward more reliable regions. Specifically, we propose the Competitive Slot Distillation (CoSDi) framework, which uses only 20 dynamically updatable learnable prototypes ($i.e.$, slots) that interact with scene features to form an affinity matrix, and imposes a zero-sum assignment along the slot dimension to drive the slots to specialize in complementary structures. Since the competition itself reveals which regions are most strongly absorbed by the slots, this affinity matrix can be directly reused as an intrinsic filtering signal to suppress unreliable regions without any auxiliary model ($e.g.$, segmentation), allowing subsequent competition to refocus on a cleaner feature pool and forming a self-reinforcing competition-filtering-refocusing loop. Experiments on eight VPR benchmarks show that CoSDi outperforms state-of-the-art methods, with a $+4.3\%$ R@1 increase on Nordland, a structurally sparse dataset with drastic seasonal changes. Moreover, without task-specific modification, CoSDi transfers directly to Cross-View Geo-Localization (CVGL), achieving a $+6.09\%$ R@1 gain over MEAN on SUES-200 Drone $\to$ Satellite cross-dataset generalization, demonstrating transferability beyond its original street-level VPR setting.