GPU-based nearest neighbor search

  1. RangeNet++, IROS 2019, with code: https://github.com/PRBonn/lidar-bonnetal, cited by 168

    Untitled

  2. point cloud ⇒ range image 64*1024 or 2048, record point 2 pixel relations

    each pixel has 5 values: range R, x, y, z, remission

    Untitled

  3. image segmentation

    assign labels back to 3d points based on the recorded relations

  4. post-processing

    1. draw-backs via range image

      1. “bleeding”: encoder-decoder hour-glass-like CNNs provide blurry outputs during inference

        1. so the border of network output, where labels changed, should be refined, by conditional random field, for example

      2. “shadowing”: Multi points (2 or more) will be projected into 1 pixel when generating range image especially when small image is used. The following images are front views, and the viewports are a little different for seeing the shadows better. 红色的两块区域对应在image中是相邻像素，而其中的的incorrect labeling，其实是"bleeding"和"shadowing"二者共同引起的。

        Untitled

    2. point label voting via range knn

      1. use absolute range difference as the distance, but the Euclidean distance works analogously, albeit being slower to compute
      2. range distance 即用图像空间的S*S窗口选择nearst neighbors，而不是radius。

    3. why the pose processing works for 642048? it should means 1 point to 1 pixel, since the total point number is 642048.

      1. This means that our approach is especially useful to help in cases of label “bleeding” or “shadowing”

      Untitled

Untitled