SPECTRA‐TOD: Reconstruction‐Driven Residual Priors for Drone‐View Tiny Object Detection

ABSTRACT

Tiny object detection in drone‐view imagery is fundamentally constrained by two coupled factors: discriminative evidence collapses at the highest‐resolution pyramid level under downsampling and fusion, while clutter and compression induce spurious low‐level activations that destabilize localization. To address this, we propose SPECTRA‐TOD , a reconstruction‐driven residual prior framework that recovers tiny‐object evidence before the detection head. SPECTRA‐TOD builds an explicit end‐to‐end chain: First, HRP probes information loss by reconstructing full‐resolution images from the low‐level pyramid features, yielding loss‐sensitive discrepancies. Second, SPRA converts these discrepancies into pyramid‐aligned residual priors by combining pixel residuals with spatial high‐pass residuals, suppressing benign low‐frequency variations. Third, PCDI injects the residual prior into the highest‐resolution pyramid feature via spatial gating, channel gating and deformable modulation, producing an enhanced feature map that directly replaces the original pyramid input to the detector head. A coupled objective links reconstruction supervision, prior stability and prior‐coupled detection optimization, ensuring that the residual prior is task‐relevant and robust. Unlike reconstruction‐difference methods that mainly use discrepancy maps as auxiliary visibility cues, SPECTRA‐TOD closes the loop by converting reconstruction loss into a pyramid‐aligned optimization prior and by explicitly substituting the enhanced into the detector head. Extensive experiments on VisDrone2019 and AI‐TOD demonstrate consistent improvements over strong detectors and tiny‐object oriented baselines. In particular, SPECTRA‐TOD achieves 29.0 AP on VisDrone2019 and 24.0 AP on AI‐TOD, and yields clear gains on very tiny and tiny subsets, validating the effectiveness of converting information loss into actionable priors for drone‐view tiny object detection.