Skip to content
Book Open access

PolyphaseDet: Sampling-Aware Visual Information Aggregation for Dense Aerial Object Mining

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 54 references

Abstract

Large-scale aerial imagery provides fine-grained visual information for urban analytics, traffic monitoring, emergency response, and geospatial data mining. A central challenge is to reliably detect tiny objects in high-resolution, dense, and cluttered scenes. Existing detector pyramids commonly rely on stride-based downsampling to aggregate multi-scale features. When objects occupy only a few pixels, such reductions may discard or alias fine spatial cues and make object representations sensitive to the sampling phase. We propose PolyphaseDet, a neck-local aggregation design for sampling-aware aerial tiny-object detection. PolyphaseDet contains two components. Polyphase-Preserving Injection (PPI) preserves phase-specific stride-4 evidence through space-to-depth reindexing before injecting it into the stride-8 representation. Phase-Offset Downsampling (POD) aggregates responses from multiple sampling offsets before committing features to a coarser grid. The backbone, decoder, matching strategy, loss function, and inference procedure are kept unchanged, isolating the effect of the proposed neck aggregation design. On VisDrone2019-DET, PolyphaseDet improves a controlled baseline from 28.8 to 32.9 AP at 640 × 640 input resolution, and reaches 40.3 AP at 1280 × 1280. On AI-TOD, it achieves 28.1 AP. Ablation studies show that the gains are not simply due to forwarding P2 features, adding extra branches, or using generic blur-based downsampling. PolyphaseDet also improves prediction stability under one-pixel input translations while maintaining real-time efficiency comparable to that of the RT-DETR baseline.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.