Research map / Detection and segmentation / CNN features
Pinwheel-shaped Convolution and Scale-based Dynamic Loss for Infrared Small Target Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes pinwheel-shaped convolution (PConv) to replace early standard convolutions and a scale-based dynamic (SD) loss for BBox and mask labels in infrared small target detection, plus a new SIRST-UAVB dataset.
Problem
Standard convolutions ignore the Gaussian pixel distribution of dim infrared small targets, and existing box/mask losses (CIoU, SLS) do not account for label IoU fluctuation or varying scale sensitivity across target sizes.
Approach
A pinwheel-shaped set of asymmetric-padded 1x3/3x1 and 3x1/1x3 convolutions concatenated and fused by a 2x2 kernel to enlarge the receptive field with few parameters, and a loss that dynamically weights scale and location terms by target area relative to a maximum IRST size (81 pixels).
Key results
- On YOLOv8n-p2 with PConv(4,3) and SDB(0.5) combined, mAP50 on SIRST-UAVB rose to 94.1 from 93.2 baseline, with IRSTD-1K mAP50 at 88.9 vs 87.4.
- SDM(0.5) loss on MSHNet raised IoU on IRSTD-1K from 66.82 to 68.49 and reduced Fa from 14.20 to 9.34 (x10^-6).
- PConv(4,3) on MSHNet improved SIRST-UAVB IoU from 23.93 to 24.66 and Pd from 72.66 to 73.70 with Fa reduced from 11.37 to 10.83.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the first two standard convolutions with PConv(4,3) (ablation) | YOLOv8n-p2 · IRSTD-1K | mAP50: 87.4 -> 89.9 | Helped |
| Replace the first two standard convolutions with PConv(4,3) (ablation) | YOLOv8n-p2 · SIRST-UAVB | mAP50: 93.2 -> 93.8 | Helped |
| Use SDB loss (δ=0.3) instead of CIoU (ablation) | YOLOv8n-p2 · IRSTD-1K | mAP50: 87.4 -> 88.6 | Helped |
| Use SDB loss (δ=0.3) instead of CIoU (ablation) | YOLOv8n-p2 · SIRST-UAVB | mAP50: 93.2 -> 93.9; IoU: 94 | Mixed |
| Use SDM loss (δ=0.5) instead of SLS (ablation) | MSHNet · IRSTD-1K | IoU: 66.82 -> 68.49 | Helped |
| Use SDM loss (δ=0.5) instead of SLS (ablation) | MSHNet · SIRST-UAVB | IoU: 23.93 -> 24.97 | Helped |
| Add PConv without changing the baseline loss (ablation) | YOLOv8n-p2 · IRSTD-1K | mAP50: 87.4 -> 89.9 | Helped |
| Add SDB loss without adding PConv (ablation) | YOLOv8n-p2 · IRSTD-1K | mAP50: 87.4 -> 88.1 | Helped |
| Add PConv without changing the baseline loss (ablation) | YOLOv8n-p2 · SIRST-UAVB | mAP50: 93.2 -> 93.8 | Helped |
| Vary SDB loss δ from 0.3 to 0.7 (ablation) | YOLOv8n-p2 · IRSTD-1K | mAP50: 88.6; 88.1; 88.8 | Helped |
| Vary SDB loss δ from 0.3 to 0.7 (ablation) | YOLOv8n-p2 · SIRST-UAVB | mAP50: 93.9; 93.8; 93.9 | Mixed |
| Vary SDM loss δ; δ=0.5 is reported as the best balance (ablation) | MSHNet · IRSTD-1K | IoU: 68.17; 68.49; 67.19 | Mixed |
Surprising findings
- PConv with a larger kernel length helps on IRSTD-1K but not on SIRST-UAVB, where kernel length 3 in the second layer gave the best balance.
- The SDM loss is limited by batch-averaged target sizes, so large batch sizes weaken the per-target scale weighting.
Limitations
- SDM loss uses batch-averaged target size, so large batch sizes reduce the benefit of per-target scale weighting.
- PConv increases parameters relative to standard convolution when replacing the first two layers (about 122% for PConv(4)).
- Bird targets were excluded from mask annotations in SIRST-UAVB.
Experimental setup
Datasets: IRSTD-1K SIRST-UAVB
Benchmarks: IRSTD-1K SIRST-UAVB
Metrics: precision recall mAP50 IoU Pd Fa
Baselines: Conv GConv DSConv DRConv LSKConv DConv MixConv AKConv CIoU DIoU GIoU IoU
Builds on: SLS loss CIoU YOLOv8 MSHNet
Closest papers by meaning
- Semi-supervised Infrared Small Target Detection with Thermodynamic-Inspired Uneven Perturbation and Confidence Adaptation (AAAI 2025 · similarity 0.88)
- Rethinking Evaluation of Infrared Small Target Detection (NeurIPS 2025 · similarity 0.88)
- Rethinking Infrared Small Target Detection: A Foundation Driven Efficient Paradigm (NeurIPS 2026 · similarity 0.88)
- Infrared Small Target Detection with Scale and Location Sensitivity (CVPR 2024 · similarity 0.87)
- IRPruneDet: Efficient Infrared Small Target Detection via Wavelet Structure-Regularized Soft Channel Pruning (AAAI 2024 · similarity 0.86)
- MOCID: Motion Context and Displacement Information Learning for Moving Infrared Small Target Detection (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.