MSHC-YOLO: multiscale heterogeneous convolution for small object detection in traffic scenes
TL;DR
This work proposes MSHC-YOLO, a Multi-Scale Heterogeneous Convolution module that decomposes standard convolution into three parallel kernels with distinct shapes to capture direction-specific spatial structures at negligible parameter cost and interacts synergistically with the P2 head, validating that the shallower the feature, the more kernel shape matters.
Abstract
Small object detection in traffic monitoring suffers from a structural inefficiency in standard detectors: isotropic 3×3 convolutions treat all spatial directions uniformly, yet traffic objects exhibit strong anisotropic geometry—pedestrians are vertically elongated, vehicles are horizontally wide. We propose MSHC-YOLO, built on two complementary innovations. First, a Multi-Scale Heterogeneous Convolution (MSHC) module decomposes standard convolution into three parallel kernels with distinct shapes—square (3×3), horizontal strip (1×3), and vertical strip (3×1)—to capture direction-specific spatial structures at negligible parameter cost. Second, a P2-Guided Fine-Grained Detection Head extends the feature pyramid to stride 4, providing a 160×160 detection grid that resolves objects as small as 8×8 pixels. Critically, MSHC and the P2 head interact synergistically: heterogeneous kernels are most effective in shallow, highresolution layers where spatial detail is preserved. On VisDrone2019-DET, MSHC-YOLO achieves 37.5% mAP@0.5, outperforming YOLO11n by 4.2 points while adding only 0.21M parameters. Ablation reveals a super-linear gain— MSHC contributes +0.5 mAP in the standard P3 path but +2.0 mAP when combined with the P2 head, validating that the shallower the feature, the more kernel shape matters.