STF-Net: A unified spatio-temporal-frequency learning framework for robust small-target detection from event streams

Event cameras offer microsecond latency, high dynamic range, and minimal motion blur, but extracting reliable small-target signatures from their asynchronous, sparse event streams remains challenging, particularly in cluttered, high-motion scenes. To address this, we propose STF-Net, a novel unified learning framework that simultaneously exploits spatial, temporal, and frequency-domain cues for robust small-target detection from event streams. The framework is built upon three key innovations. First, a stage-optimized backbone progressively transitions from comprehensive Temporal-Spatial-Frequency (TSF) blocks to lightweight Spatial-Frequency (SF) blocks, achieving a favorable balance between representational capacity and computational efficiency. Second, a joint Spatio-Frequency Enhancement Module (SFEM) integrates local spatial attention with global frequency processing, employing stage-specific frequency masking to extract discriminative multi-scale features. Third, a Temporal Attention Recurrent Unit (TARU) combines temporal attention and recurrent modeling to capture both fine-grained motion and long-range dependencies, enabling holistic temporal understanding. Extensive experiments confirm the superior performance of STF-Net against seven state-of-the-art object detection methods. On the challenging NeRDD dataset, our model achieves a state-of-the-art mean mAP of 51.6% ± 0.4% across three independent runs, surpassing the RVT baseline (mean mAP: 49.9% ± 0.4%) by 1.7 percentage points. STF-Net also achieves superior mean scores in localization precision (mean AP₇₅: 55.6% ± 0.4%) and small-target recall (mean AP_S: 50.3% ± 0.5%), exceeding RVT by 2.1 and 1.8 points, respectively. Crucially, STF-Net maintains high parameter efficiency (16.1M) and delivers a real-time inference speed of 69.9 FPS, more than double the standard 30 FPS threshold, despite a moderate increase in computational latency compared to the fastest baseline. Extensive evaluations on the Gen1 and 1Mpx automotive benchmarks further demonstrate the model's strong generalization ability. This work establishes a unified spatio-temporal-frequency learning paradigm that effectively addresses the core challenges of accurate and robust perception from event streams.