Learning Where to Embed: Noise-Aware Positional Embedding for Query Retrieval in Small-Object Detection

• Cluttered aerial scenes combine tiny targets, scale variation, dense layouts, and extensive background, degrading localization and recognition.Source: §1 · p. 1
• Background responses pollute queries and force decoders to repeatedly refine low-quality embeddings, increasing computational cost.Source: §1 · p. 2
• HELP selectively preserves positional encodings in foreground-salient regions while suppressing them in background-dominant regions.Source: §1 · p. 2
• HPE provides heatbar visualization for interpretable analysis and fine-tuning of selective positional embedding.Source: §1 · p. 2
• MOHFE and HQ-Retrieval provide noise-suppressed encoding and gradient-based query filtering, reducing decoder depth from eight to three.Source: §1 · p. 2
• 59.4% parameter reduction (66.3M vs. 163M) is achieved by integrating LSConv while accelerating training and inference.Source: §1 · p. 2
• The approach is validated on five benchmarks with consistent gains and good scalability.Source: §1 · p. 2

• CNN-based detectors use hierarchical feature extraction and multi-scale representations, with one-stage methods emphasizing efficiency and two-stage methods using region proposals.Source: §2.1 · p. 2
• Transformer detectors use set prediction and query-driven decoding with global context modeling, often combining convolutional backbones, transformer encoders, and learnable queries.Source: §2.2 · p. 2
• HELP differs in the supplied overview by combining LSConv, heatmap-guided encoder fusion, and high-quality query retrieval within one pipeline.Source: Figure 1 · p. 3
• Efficiency-oriented transformer methods include dynamic token pruning and allocation, while other small-object strategies use focal loss, class activation maps, and sparse attention.Source: §2.2 · p. 3

Heatmap-guided Positional Embedding

• HELP uses heatmap-guided positional embedding to preserve positional cues in foreground-salient regions and suppress them in background-dominant regions.The heatmap-guided embedding is designed to align positional information with detection semantics.Source: §3.1 · pp. 3–4
• Second- and third-order gradients of classification confidence produce class-discriminative heatmaps, while regression-based backpropagation produces geometry-aware heatmaps.The semantic and geometric heatmaps are combined into a mixed heatmap for subsequent masking.Source: §3.1 · pp. 3–4
• HPE combines class and bounding-box heatmaps into a mixed heatmap, then thresholds it into a binary mask for positional encoding.Locations with Hmap ≤ τ suppress positional encodings; locations with Hmap > τ preserve them.Source: §3.1 · p. 4
• Heatmap supervision is used only during training, so inference uses a standard forward pass without additional gradient computation.The learned heatmap-driven embedding behavior remains in the model parameters after training.Source: §3.1 · p. 4

Encoder–Decoder Integration

• MOHFE fuses class-driven and box-driven embeddings into encoder features, while HQ-Retrieval derives compact decoder queries from mixed-heatmap embeddings.These components provide cleaner encoder memory and prioritize foreground-salient positions before deformable-attention decoding.Source: §3.2 · p. 4, §3.2.1 · pp. 4–6, §3.2.2 · p. 6

Linear-Snake Convolution

• LSConv enriches retrieval-relevant features with geometry-aware sampling before heatmap-guided query retrieval.Its complementary sampling paths support sparse, complex small-object representations.Source: §3.3 · p. 6

Datasets

• Five public benchmarks cover generic and aerial detection, including NWPU VHR-10, PASCAL VOC, DOTA, DIOR, and VisDrone.The datasets assess scalability across data types, object densities, and scene complexity.Source: §4.1 · p. 7

Experimental protocol

• Official training/testing splits and identical hyperparameters are used across datasets, with λ = 0.5 and τ = 0.5 selected on validation data.The heatmap threshold τ binarizes the heatmap for query retrieval.Source: §4.1 · p. 7

Comparisons and ablations

• Table 1 compares the method with RT-DETR on PASCAL VOC and NWPU VHR-10, while Table 2 evaluates component contributions on both datasets.Source: Table 1 · p. 8, Table 2 · p. 8

Metrics

• Accuracy is reported with mAP@50 and mAP, while GFLOPs and Params (M) measure computational cost and model size.Source: §4.1 · p. 8

Quantitative results

• 94.51% mAP@0.5 and 67.20% mAP are achieved on NWPU VHR-10.Source: §4.2.1 · p. 9
• +1.91 mAP@0.5 and +6.94 mAP are obtained over RT-DETR, while computation falls from 136 to 57 GFLOPs.Source: §4.2.1 · p. 9
• 59.4% fewer parameters are used than RT-DETR, reducing model size from 163M to 66.3M.Source: §4.2.1 · p. 9

HPE visualization

• HPE heatbars exhibit a hot-center, cold-ends profile that emphasizes object-relevant regions and suppresses background responses during query initialization.Source: §4.2.2 · p. 9, Figure 6 · p. 9

Component-wise ablation

• The full configuration achieves the best performance on both NWPU VHR-10 and PASCAL VOC, indicating complementary component gains.Source: §4.3.1 · p. 9
• 94.51% mAP@0.5 and 67.20% mAP are reached on NWPU VHR-10 after adding HQ-Retrieval.Source: §4.3.1 · p. 10
• The full model improves PASCAL VOC mAP@0.5 from 69.4% to 70.6% and mAP from 50.5% to 51.6%.Source: §4.3.1 · p. 10

Decoder depth

• At 100 epochs, Det2 reaches peak performance at 70.5% AP, whereas deeper decoders help under limited training budgets.Source: §4.3.2 · p. 10
• Three decoder layers provide a favorable trade-off in data-scarce regimes, while GFLOPs increase modestly from 53.7 to 57.0 across depths.Source: §4.3.2 · p. 10

Additional benchmarks

• Relative to DETR, improvements of +6.4/+3.9 on DOTA, +7.6/+5.4 on DIOR, and +7.5/+4.4 on VisDrone are reported for AP50/AP75.Source: §4.3.3 · pp. 10–11
• The method consistently outperforms DETR across DOTA, DIOR, and VisDrone at both AP50 and AP75.Source: §4.3.3 · pp. 10–11

Conclusion

• HELP preserves positional embeddings in foreground-salient regions and suppresses them in background-dominant regions.Source: §5 · p. 11
• Selective positional embedding produces cleaner decoder inputs, improves high-IoU localization, and reduces reliance on deep decoder stacks.Source: §5 · p. 11
• The paper identifies embedding placement as a noise-aware allocation problem for improving detector accuracy and efficiency.Source: §5 · p. 11