Mask RCNN has very magical function, ability of target detection and image segmentation at pixel level, recently, with the model experiment, the long and thin target detection effect is not ideal, especially the tilt of the target, can be considered as overlapping and filtered, there may be is a problem of the maximum inhibition may also be the anchor setting problem, effect, as shown in the figure to everybody together discuss,

CodePudding user response:

Using threshold change and the rotation Angle of the segmentation results are as follows, still not ideal:

CodePudding user response:

Should be test box is consistent with goals from vertical edge was right, as shown in figure on the right side of the pile of text:

CodePudding user response:

Test box is transformed from vertical edges are consistent to fish, you can write a program to realize testing box labels transformation, is tilted from vertical box to box, in this way, the original image of target segmentation and detection box is tilted, and then to training network model,

CodePudding user response:

White teacher, if a large figure, the resolution of the width of more than 4000, to identify multiple targets, densely packed inside can lay figure into 1024 training within first, then take the picture out prediction?

CodePudding user response:

Friendly to elongated objects not is true,

CodePudding user response:

Is the problem of anchor, each time will only generate square, and tilt your generated in the form of the candidate box greatly again, the correct ratio is very small, directly filtered out

CodePudding user response:

Consider modifying the RPN generation how to generate tilt of the anchor

CodePudding user response:

https://zhuanlan.zhihu.com/p/108255937

CodePudding user response:

You use what CNN, traditional method is simple, know about the license plate recognition system of segmentation algorithm of the characters

CodePudding user response:

refer to 12 floor TR0217 reply:

you what CNN, this traditional method is simple, look at the license plate recognition system of segmentation algorithm of character

To fix the blame, so many uneven area, to binarization impossible

CodePudding user response:

Topic main now have any other comments, I also doing narrow object detection, have some problems, could you ask