from Deep Learning Based Semantic Labelling of 3D Point Cloud in Visual SLAM

1. 超體素方法進行預分割，將點雲根據相似性變成表層面片（surface patches）降低計算複雜度。

         將場景分割問題轉換為圖分割問題（graph partitioning problem）

• Method 1：Mean-shift聚類演算法 計算node之間的距離

1. node指的是每個patch，連線node之間的line就是相鄰patch的共邊；
2. 距離可以是歐氏距離，也可以是馬氏距離；
3. Mean-shift演算法可見簡單介紹及Python實現或者簡單的機器學習演算法Mean-shift演算法

缺點：計算量太大

• Method 2：利用面片的法向量方法  聚類  法向量可以表示出區域性凸性資訊。

缺點：當noise太多的時候可靠性降低。

• 最終使用method 2 結合可靠性平面來做分割 最後使用圖割法分割

關於2D Object Detection and Semantic Segmentation

An essential component to get semantic information is object detection

, which can localize object instances in images. Girshick et al. [21] presented R-CNN, which proposed to apply CNN to object detection. Other similar methods have been proposed in recent years, like Fast R-CNN [22], Faster RCNN [23], Mask-RCNN [24] and YOLO [25-26]. R-CNN uses selective search algorithm for generating region proposals, which runs very slow. Faster R-CNN replaces the slow selective search algorithm with a fast neural net. Mask R-CNN improves the region of interest (ROI) pooling layer and extends Faster R-CNN to pixel-level image segmentation
Semantic segmentation is to understand an image at a pixel level, which can label each pixel with a class identity. Similar to object detection, state-of-the-art semantic segmentation approaches also rely on CNN. FCN [4] by Long et al. is the first end-to-end system, which popularizes CNN architecture for semantic segmentation. U-Net [5] is a popular encoder-decoder architecture which can make use of annotated samples more efficiently and have a higher accuracy. SegNet [6] is a similar encoderdecoder architecture. SegNet copies indices from max-pooling for up-sampling, which makes it more memory efficient. RefineNet [7] proposes a method called RefineNet block which fuses both high resolution and low resolution features. It solves the problem of significant decrease in image resolution when we repeat the sub-sampling operation. PSPNet [8] introduces a pyramid pooling method to aggregate the context. DeepLab [9-11] utilizes dilated convolutions to increase the field of view.