GeoGAT: Bidirectional Temporal Sampling Meets Hierarchical Graph Attention for Global Video Geo-localization

1 Introduction

• Earlier retrieval-based methods focus on predefined local regions and do not scale to global coverage.Source: §1 · p. 1
• Existing methods use one-way uniform sampling and independent hierarchy classifiers, despite sparse geographic cues and strict city-to-continent inclusion constraints.The paper says inference-only consistency constraints leave hierarchical conflicts unresolved during learning.Source: §1 · pp. 1–2
• GeoGAT couples forward and offset-reversed frame sampling with graph attention over a geographic hierarchy and dual constraints during training and inference.The dual constraints combine soft parent-child correlation during training with hard geographic validity enforcement during inference.Source: §1 · p. 2
• GeoGAT10k comprises 9,720 multi-shot edited videos from 166 cities worldwide to evaluate generalization on complex video structures.Source: §1 · p. 2
• In extensive experiments on CityGuessr68k and GeoGAT10k, GeoGAT achieves state-of-the-art performance across all four geographic hierarchies.Source: §1 · p. 2

2.1 Paradigms of Video Geo-localization

• Global classification offers an alternative to retrieval’s reference-database dependence, which makes global scaling impractical.Source: §2.1 · p. 2

2.2 Retrieval-based Region-level Paradigm

• Retrieval-based methods match video frames against satellite-image databases and are effective in local regions, but database costs rise with region size and applicability remains limited to predefined areas.Source: §2.2 · p. 2

2.3 Classification-based Global Paradigm

• CityGuessr formulates global video geo-localization as classification with four independent linear heads, and its CityGuessr68k release is the first large-scale global video geo-localization benchmark.Source: §2.3 · p. 3
• GeoGAT10k is constructed to evaluate generalization under complex video structures and provide richer data support.Source: §2.3 · p. 3

3.1 Problem Formalization

• The task is to preserve geographic information in a video representation while predicting its city, state, country, and continent labels.Only some frames contain discriminative cues, while others carry location-irrelevant dynamics.Source: §3.1 · p. 3

3.2 Bidirectional Temporal Sampling

• Bidirectional sampling combines forward and offset-reversed frame sequences, whose disjoint temporal coverage can expose complementary geographic evidence.Both sequences are encoded and their features concatenated and projected into a joint representation.Source: §3.2 · p. 3

3.3 GAT-based Geographic Hierarchy Classifier

• GeoGAT models geographic labels as a directed hierarchy graph and uses GAT message passing instead of independent classifiers.The graph encodes city-to-continent parent-child relationships; soft and hard constraints enforce consistency in logits and inference.Source: §3.3 · pp. 3–4, Figure 2 · p. 4

3.4 Multi-Granularity Loss

• A multi-granularity loss combines mutual-information maximization with direct supervision of predictions at each hierarchy.The joint objective is L = LGAT + λ · LMI, with λ balancing the two terms.Source: §3.4 · pp. 4–5

4.1 Dataset and Metrics

• GeoGAT10k is a TikTok-sourced collection of 9,720 multi-shot edited videos spanning 166 cities; CityGuessr68k has 54,614 training and 13,655 test videos under an 80:20 split.CityGuessr68k's stratified split preserves class distributions across city, state, country, and continent.Source: §4.1 · p. 5

4.2 Training Details

• The evaluation uses Top-1 accuracy for city, state/province, country, and continent; GeoGAT uses bidirectional sampling of 16 frames and a two-layer GATv2.GeoGAT uses VideoMAE-Small pretrained on Kinetics-400; training runs for 10 epochs on CityGuessr68k and 20 on GeoGAT10k.Source: §4.1 · p. 5, §4.2 · p. 5

4.3 Performance on CityGuessr68k

• GeoGAT achieves the best performance at every granularity against both VidTAG variants on CityGuessr68k.VidTAG-ret‡ outperforms its classifier-adapted counterpart, but GeoGAT still surpasses it.Source: §4.3 · p. 6, Table 1 · p. 6

4.4 Performance on GeoGAT10k

• GeoGAT beats CityGuessr by 24.3, 24.3, 28.4, and 26.4 percentage points at the city, state, country, and continent levels on GeoGAT10k.All methods have lower accuracy on GeoGAT10k than on CityGuessr68k; the paper describes GeoGAT10k as harder because it contains abrupt scene transitions and visual effects.Source: Figure 5 · p. 6, §4.4 · p. 6

4.5 Ablation Study

• Across both datasets, graph heads beat linear heads, BTS beats OUS and BRS, hard constraints improve all hierarchies, and LMI improves results.Soft parent-child coupling can degrade fine-grained performance; the full model attains the best performance across all metrics.Source: §4.5 · p. 6

4.6 Attention Flow Analysis

• Attention is strongest on each hierarchy's own features, with cross-hierarchy flow concentrated between adjacent levels and diminishing with depth.The authors connect locality to redundancy and over-smoothing from soft coupling, and say hard validity enforcement compensates for a continent-level information bottleneck.Source: Figure 6 · p. 7, §4.6 · p. 7

4.7 Performance on Sparse Sample Regions

• On the 38-city sparse-sample subset, GeoGAT performs best across hierarchies, with its advantage widening from city to continent.Without constraints, GeoGAT reduces conflict rates to 7.8%, versus 17.8% for CityGuessr and 19.1% for VideoMAE; both constraints eliminate conflicts entirely.Source: §4.7 · p. 7, Figure 7 · p. 7

• GeoGAT achieves state-of-the-art accuracy across four geographic hierarchies on both CityGuessr68k and GeoGAT10k.Its gains stem from complementary temporal views and structured geographic reasoning.Source: §5 · p. 7
• 2.6 percentage points at the city level: GeoGAT outperforms strongest baseline VidTAG under both classification and retrieval protocols.Source: §5 · p. 7
• GeoGAT generalizes robustly to multi-shot edited videos, where existing methods degrade substantially.Source: §5 · p. 7
• GeoGAT10k contains 9,720 videos from 166 cities featuring real-world editing patterns.Source: §5 · p. 7