<p>Accurate parcel segmentation is critical for real-time logistics automation, particularly under complex conditions such as occlusion and dense stacking. While deep learning-based detectors have shown promise, conventional bounding box methods often fail to provide precise localization required by robotic systems. To address this, we propose CheapConv, a lightweight multi-scale convolutional operator specifically designed for efficient dense detection and segmentation in logistics scenarios. We integrate CheapConv into a re-parameterized YOLOv8 framework, called YOLOv8-cheap, which is a purely convolutional network without attention. The proposed architecture combines dense, sparse, and decomposed convolutions in a multi-branch design, enabling rich spatial feature extraction at low computational cost. Additionally, we introduce RepDS and RepHead, which leverage structural re-parameterization to improve inference-time performance without increasing complexity. Extensive experiments on the Live Stacked Carton Dataset (LSCD) and SKU-110K benchmark validate the effectiveness of our design. YOLOv8-cheap achieves notable gains in segmentation accuracy with significantly fewer FLOPs and parameters. YOLOv8n-cheap improves <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1779_Article_IEq1.gif" Format="GIF" Height="21" Rendition="HTML" Resolution="72" Type="Linedraw" Width="60" /> </InlineMediaObject> <EquationSource Format="TEX">\(AP^{{Mask}}_{50:95}\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msubsup> <mi>P</mi> <mrow> <mn>50</mn> <mo>:</mo> <mn>95</mn> </mrow> <mrow> <mi mathvariant="italic">Mask</mi> </mrow> </msubsup> </mrow> </math></EquationSource> </InlineEquation> by 1.8% while reducing computational cost by 20%. Our model also achieves 60.9% <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1779_Article_IEq2.gif" Format="GIF" Height="21" Rendition="HTML" Resolution="72" Type="Linedraw" Width="52" /> </InlineMediaObject> <EquationSource Format="TEX">\(AP^{{Box}}_{50:95}\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msubsup> <mi>P</mi> <mrow> <mn>50</mn> <mo>:</mo> <mn>95</mn> </mrow> <mrow> <mi mathvariant="italic">Box</mi> </mrow> </msubsup> </mrow> </math></EquationSource> </InlineEquation> on SKU-110K, setting a new state-of-the-art for dense object detection. The proposed approach offers a practical and scalable solution for vision-based parcel handling. Code is available at: <a href="https://github.com/li1-maker/YOLOv8-cheap">https://github.com/li1-maker/YOLOv8-cheap</a>.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Cheapconv: efficient parcel segmentation and detection for real-time logistics automation

  • Bingqian Zou,
  • Wen Li,
  • Zhengjia Chen

摘要

Accurate parcel segmentation is critical for real-time logistics automation, particularly under complex conditions such as occlusion and dense stacking. While deep learning-based detectors have shown promise, conventional bounding box methods often fail to provide precise localization required by robotic systems. To address this, we propose CheapConv, a lightweight multi-scale convolutional operator specifically designed for efficient dense detection and segmentation in logistics scenarios. We integrate CheapConv into a re-parameterized YOLOv8 framework, called YOLOv8-cheap, which is a purely convolutional network without attention. The proposed architecture combines dense, sparse, and decomposed convolutions in a multi-branch design, enabling rich spatial feature extraction at low computational cost. Additionally, we introduce RepDS and RepHead, which leverage structural re-parameterization to improve inference-time performance without increasing complexity. Extensive experiments on the Live Stacked Carton Dataset (LSCD) and SKU-110K benchmark validate the effectiveness of our design. YOLOv8-cheap achieves notable gains in segmentation accuracy with significantly fewer FLOPs and parameters. YOLOv8n-cheap improves \(AP^{{Mask}}_{50:95}\) A P 50 : 95 Mask by 1.8% while reducing computational cost by 20%. Our model also achieves 60.9% \(AP^{{Box}}_{50:95}\) A P 50 : 95 Box on SKU-110K, setting a new state-of-the-art for dense object detection. The proposed approach offers a practical and scalable solution for vision-based parcel handling. Code is available at: https://github.com/li1-maker/YOLOv8-cheap.