segment result

[qxdaaaaa]

I have reproduced your code! you are great!But I can't get your result.as you show in video,i can't get the segment picture,in other words,the result is not good.and the dataset you use is small,my is bigger,so you deal with it?
Thank you.my email is 1820093354@qq.com.

[qxdaaaaa]

@qixuxiang

[qixuxiang]

Actually, it is normal that you cannot reproduce results as I show in the video with other datesets. Because when you use more images or bigger datasets, you will get a bigger map, which is difficult to segment them into perfect pieces or voxels.

The project is not perfect, it works well only in small indoor RGBD sequences.

[qxdaaaaa]

@qixuxiang Can you share your own dataset?and In fact,I have make my own lab-environmen dataset,but the result is not good,How can i modify the parameter to get a better results?Can you give me some advice?
Thank you。

[qixuxiang]

I test my code on tum dataset, I forget which RGBD sequences I used, you can download some RGBD sequences to test.

[qxdaaaaa]

I have fund the tum dataset that you used.But the dataset is so bogger,I can't get the good result as you can.How did you deal with it?@qixuxiang

[qixuxiang]

ORB-SLAM2 has its own key-frame strategy. In my code, I use yolo v3 to process keyframe images to get semantic information and project them to 3D dense point cloud map. However, you can never get same keyframe images in different time.

My suggestion is you can try my dataset for sevaral times, or you can change params in src/segmentation_test.cc:

 config.noise_threshold = 0.01; // 1cm
 config.voxel_resolution = 0.008f;// 0.8cm
 config.seed_resolution = 0.08f; // 8cm
 config.min_plane_area = 0.01f; // m^2;
 config.max_curvature = 0.01;