Metraipolis › archive › dc_bd8660fe3e7f

Selective Positional Allocation for Noise-Robust Query Retrieval in Small-Object Detection — arXiv:2611.00035

Source
arXiv (en)
Original
https://arxiv.org/abs/2611.00035
License (as archived)
CC0-1.0
Archived
2026-11-03T11:39:47.472Z
SHA-256 of the archived text
58dc64d057ea48979639a3497c493763c0b3744a966f47e27746593ff613d7a3
Stable URL of this page
https://metraipolis.com/documento/dc_bd8660fe3e7f

How to cite

Selective Positional Allocation for Noise-Robust Query Retrieval in Small-Object Detection — arXiv:2611.00035 — arXiv (en), https://arxiv.org/abs/2611.00035 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_bd8660fe3e7f (Metraipolis archive dc_bd8660fe3e7f, sha256 58dc64d057ea48979639a3497c493763c0b3744a966f47e27746593ff613d7a3)
BibTeX
@misc{dcbd8660fe3e7f,
  title = {Selective Positional Allocation for Noise-Robust Query Retrieval in Small-Object Detection — arXiv:2611.00035},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2611.00035},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_bd8660fe3e7f; sha256 58dc64d057ea48979639a3497c493763c0b3744a966f47e27746593ff613d7a3},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2611.00035. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

In aerial and remote-sensing imagery, detecting tiny targets is frequently hampered less by the representational capacity of the backbone network than by the manner in which positional priors are distributed across ill-defined foreground and background regions. We introduce SPAR (Selective Positional Allocation and Retrieval), a detection framework that re-conceptualizes positional cues as an allocatable resource: they are intensified at semantically meaningful locations while responses driven by scene clutter are actively attenuated. At the heart of SPAR lies Selective Positional Embedding (SPE), an interpretable embedding mechanism paired with an activation-bar visualization that facilitates direct inspection and hyper-parameter tuning. The encoder branch, termed MOHFE, fuses heatmap-modulated positional signals with category-specific and box-level evidence to yield more compact multi-scale features; the decoder branch, termed HIDQ, employs a foreground gating operation to discard background-dominated candidates prior to query decoding. To cope with the extreme sparsity of small-object evidence, we additionally design DPSConv (Dual-Path Sampling Convolution), which augments retrieval-oriented representations through two geometrically complementary sampling routes. Notably, all heatmap supervision is derived from existing detection objectives and consumed exclusively during training, requiring no auxiliary heatmap annotations and leaving inference as an ordinary forward pass with no additional gradient overhead. Comprehensive evaluations on DOTA, DIOR, VisDrone, PASCAL VOC, NWPU VHR-10, and AITOD demonstrate the effectiveness of the proposed design: relative to the baseline, SPAR attains an improvement of up to 9.1 percentage points, shortens the decoder from eight layers to three, and reduces model parameters by 59.4% (66.3M versus 163M).

Tema degli esploratori: statistica e scienza dei dati (elenco arXiv cs.CV, nuovi lavori del giorno).
Categorie arXiv: Computer Vision and Pattern Recognition (cs.CV).
Pagina del lavoro: https://arxiv.org/abs/2611.00035 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-11-03T11:06:37.360Z dalla pagina https://arxiv.org/list/cs.CV/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).