How Dynamic Snake and Pinwheel Convolutions Boost Small‑Target Segmentation Accuracy

Paper 1: Dynamic Snake Convolution based on Topological Geometric Constraints for Tubular Structure Segmentation

Method

Dynamic Snake Convolution (DSConv) : an iterative offset strategy that lets the convolution kernel adaptively focus on slender, curved regions, preventing receptive‑field drift common in deformable convolutions.

Multi‑view Feature Fusion : DSConv generates several morphological kernel templates that observe the structure from different angles; random dropping suppresses redundant noise.

Topological Continuity Constraint Loss (TCLoss) : a loss built on Persistent Homology that enforces topological continuity of the segmentation, reducing fractures.

Innovations

Geometric Prior Injection : first to embed tubular snake‑shaped morphological priors into convolution kernels, markedly improving perception of thin‑walled structures.

Topological Constraint Optimization : quantifies connectivity and loops via persistent homology to guide the network toward fracture regions, outperforming classic losses such as clDice.

Cross‑dimensional Generality : applicable to both 2D (retinal vessels, road networks) and 3D (coronary arteries) data.

Experimental Results

Volume Accuracy : Dice coefficient improves by 1.3%–3.4% and clDice by 0.8%–3.8% on DRIVE, Massachusetts Roads, and Cardiac CCTA datasets.

Topological Continuity : β0/β1 errors drop by 0.2–0.3 and the Overlap‑to‑First‑Error (OF) metric rises by 3.3%–6.0%.

Visualization : DSConv repairs vessel breaks that traditional methods cause and focuses more precisely on tubular regions.

Paper link: https://arxiv.org/pdf/2307.08388

Paper 2: Pinwheel‑shaped Convolution and Scale‑based Dynamic Loss for Infrared Small Target Detection

Method

Pinwheel Convolution (PConv) : designs an asymmetric kernel based on the Gaussian spatial distribution of infrared small‑target (IRST) pixels; expands receptive field by ~177% with only ~22% parameter increase.

Scale‑Dynamic Loss (SD Loss) : dynamically adjusts the weights of IoU loss (Sloss) and position loss (Lloss) according to target area, mitigating large IoU fluctuations for tiny targets.

New Benchmark SIRST‑UAVB : a UAV/air‑bird dataset containing 3,000 images, smallest targets of 9 pixels, with complex backgrounds (clouds, buildings, etc.).

Innovations

Data‑driven Convolution Design : encodes the Gaussian imaging statistics of IRST into the convolution shape, replacing generic kernels such as the first layer of YOLOv8.

Adaptive Loss Function : introduces dynamic coefficients β_B/β_M that adjust loss weights in real time based on target scale, outperforming static losses like CIoU and NWD.

Small‑Target Optimization : on SIRST‑UAVB, PConv + SD Loss raises mAP₅₀ from 87.4% to 93.8% and reduces false‑alarm rate to 10.83 × 10⁻⁶.

Experimental Results

Detection Task : YOLOv8n‑p2 + PConv(4,3) + SDB(δ=0.5) achieves mAP₅₀ = 93.8%, surpassing MixConv, AKConv and other variants.

Segmentation Task : MSHNet + PConv + SDM raises IoU from 66.82% to 68.49%; accuracy for targets < 16 pixels improves markedly.

Ablation Study : the optimal PConv blade size (4,3) combined with δ = 0.5 in SD Loss validates their synergistic effect.

Paper link: https://arxiv.org/abs/2412.16986

Summary and Future Directions

Dynamic convolutions that embed geometric or statistical priors (e.g., tubular snake shapes, IRST Gaussian distribution) together with adaptive mechanisms (topological constraints, scale‑dynamic loss) substantially improve segmentation and detection of thin structures or tiny targets. Future work may extend these kernels to other domains such as neural‑fiber tracking in medical imaging and integrate Neural Architecture Search to automatically optimise convolution shapes and loss weights.

Code example

收
藏
，
分
享
、
在
看
，
给
个
三
连
击呗！