Optics and Precision Engineering, Volume. 32, Issue 20, 2987(2024)

Shape from polarization based on sparse self-attention

Zhichao YU1... Zhenhua WAN2 and Kaichun ZHAO1,* |Show fewer author(s)
Author Affiliations
  • 1Department of Precision Instrument, Tsinghua University, Beijing00084, China
  • 2Faculty of Mechanical Engineering, Guangxi University, Nanning530004; China
  • show less
    Figures & Tables(12)
    Schematic diagram of reflective light polarization measurement
    CMOS of polarization camera
    Schematic diagram of attention calculation
    Schematic diagram of bi-level routing sparse attention calculation
    Overview of the proposed model
    Results of selected test samples from the DeepSfP dataset
    Results of different methods on our test data
    • Table 1. Quantitative evaluation results of the DeepSfP dataset

      View table
      View in Article

      Table 1. Quantitative evaluation results of the DeepSfP dataset

      方法按图像按像素
      平均误差SSIM均值中值均方根11.25°22.5°30.0°
      Smith et al.36.240.87436.2931.8143.0812.032.246.7
      DeepSfP19.160.90018.5714.8423.3335.871.983.3
      sfp-wild14.560.92414.0210.0319.2255.383.190.3
      本文方法13.370.93012.798.7317.9860.584.891.3
    • Table 2. Quantitative evaluation of different methods on our test data

      View table
      View in Article

      Table 2. Quantitative evaluation of different methods on our test data

      方法角度误差精度
      均值中值均方根误差11.25°22.5°30.0°
      Smith et al.34.4933.2438.196.425.042.4
      sfp-wild22.8519.4927.2121.758.674.7
      本文方法18.2615.1522.5334.872.485.7
    • Table 3. Results of the number of self-attention modules on performance

      View table
      View in Article

      Table 3. Results of the number of self-attention modules on performance

      自注意力模块数量按图像按像素
      平均误差SSIM均值中值均方根11.25°22.5°30.0°
      213.370.93012.798.7317.9860.584.891.3
      413.530.93012.928.7218.2060.584.491.0
      813.830.92913.199.0618.3459.283.890.7
    • Table 4. Results of different polarization representation methods on performance

      View table
      View in Article

      Table 4. Results of different polarization representation methods on performance

      偏振表示方式按图像按像素
      平均误差SSIM均值中值均方根11.25°22.5°30.0°
      本文方式13.370.93012.798.7317.9860.584.891.3
      Sfp-wild13.890.92813.199.0118.4559.483.790.3
    • Table 5. Results of different attention modules

      View table
      View in Article

      Table 5. Results of different attention modules

      注意力类型按图像按像素
      平均误差平均处理时间均值中值均方根11.25°22.5°30.0°
      本文方式13.370.54812.798.7317.9860.584.891.3
      多头注意力13.540.56212.888.8218.1460.584.590.9
    Tools

    Get Citation

    Copy Citation Text

    Zhichao YU, Zhenhua WAN, Kaichun ZHAO. Shape from polarization based on sparse self-attention[J]. Optics and Precision Engineering, 2024, 32(20): 2987

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category:

    Received: May. 9, 2024

    Accepted: --

    Published Online: Jan. 10, 2025

    The Author Email: ZHAO Kaichun (kaichunz@mail.tsinghua.edu.cn)

    DOI:10.37188/OPE.20243220.2987

    Topics