Some questions about self-made data processing

[blandness1217]

Hello, your work is outstanding, and I want to use your network to process the data I have collected. As you can see, there are two tubes of the same shape in Figure1. Figure 2 is a point cloud captured with a depth camera, then used Cloudcompare to segment the background and add instance labels for each tube. Finally, we can get the point cloud data in .txt format, which contains [x, y, z, instance_label]. Each tube is composed of about more than 2000 points. I know that only 2 elements are too few, and I plan to buy more tubes to increase it to 5 later. So here are my questions: Q1: Do you think using your network to train my data can achieve the purpose of instance segmentation? Q2: I noticed that there is a ‘center score’ in the data composition used for training. How can I add ‘center score’ for each point in my data? Q3: How can I determine the parameters during training? ‘d_max’ for example. Looking forward to your soon reply.

[xyjbaal]

Thank you for your attention to our work. A2: If you have collected your dataset, please refer to this https://github.com/waiyc/Bin-Picking-Dataset-Generation.git. The center score is calculated in step 2. A3: d_max is the largest size of your model (tube).

[blandness1217]

I am appreciate for your help！ I will try to process my data right now and please keep this issue open to enable me to contact with you later.

[blandness1217]

Hello, I got the _c_map.txt file after running fpcc_test. The content is [x, y, z, center_score]. How can I get the same picture as in Figure 6(b) of the article?

[xyjbaal]

You can visualize the center score through CloudCompare, with XYZ being the coordinates and treating the scores as an intensity value.

[blandness1217]

Thanks!! I got it.

[blandness1217]

These days I tried to make my own dataset and use fpcc to train, and finally got satisfactory results. I sincerely thank you for your help and wish you well.

[xyjbaal]

It's really nice to see such a result!

[mai4567]

These days I tried to make my own dataset and use fpcc to train, and finally got satisfactory results. I sincerely thank you for your help and wish you well.

hello! I am interested in your result. Can you get the good result if the tubes are heavly occuled by each other?

[blandness1217]

Due to financial constraints, I only purchased 6 parts, so I did not build a severely stacked scene. The instance segmentation results of some partially occluded scenes are shown in the figure below.

[mai4567]

Due to financial constraints, I only purchased 6 parts, so I did not build a severely stacked scene. The instance segmentation results of some partially occluded scenes are shown in the figure below.

hello! I just ran the fpcc_test.py by XA dataset. And I found that the output file like this. the RGB channels seem to be wrong. Then I debug the fpcc_test.py and I found the ins_pre seem to be wrong. Can you give me some advices?

[blandness1217]

Try using cloudcompare to visualize it. The first three columns are coordinates, and the last three columns are RGB (0-1)

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2024年5月27日(星期一) 下午4:25 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [xyjbaal/fpcc_pytorch] Some questions about self-made data processing (Issue #5)

Due to financial constraints, I only purchased 6 parts, so I did not build a severely stacked scene. The instance segmentation results of some partially occluded scenes are shown in the figure below.

hello! I just ran the fpcc_test.py by XA dataset. And I found that the output file like this. image.png (view on web) the RGB channels seem to be wrong. Then I debug the fpcc_test.py and I found the ins_pre seem to be wrong. Can you give me some advices? 1716798207581.png (view on web)

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>