search

WeijingShi / Point-GNN

Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud, CVPR 2020.
MIT License
523 stars 114 forks source link

自己采的数据集检测不到目标，没有框 #64

Open GXQ0527 opened 3 years ago

GXQ0527 commented 3 years ago

@WeijingShi 作者，您好！我训练网络之后，换成自己采集的数据集检测不到目标，但是kitti的数据集能够检测到目标，不知是什么原因，希望作者能给予答案！谢谢！

WeijingShi commented 3 years ago

Hi @GXQ0527, Did you train the model with your custom data or the Kitti data?

If you trained it on Kitti and it works for Kitti but not your custom data, then there is a gap between your custom data and Kitti. A common issue is on the coordinate system. Make sure your raw lidar and calibration file can provide points in the Kitti camera coordinates (y-axis is the height direction pointing to the ground). The following function is in charge of reading the point cloud: https://github.com/WeijingShi/Point-GNN/blob/48f3d79d5b101d3a4b8439ba74c92fcad4f7cab0/dataset/kitti_dataset.py#L666 Hope it helps.

GXQ0527 commented 3 years ago

y轴是向下的，我用的是速腾128线的激光雷达，而kitti用的是velodyne64线的，不知道是不是因为这个原因，我检测时open3d效果很好，但是没有检测框，二维图像上也没有检测框。

WeijingShi commented 3 years ago

Hi @GXQ0527, Sorry I have not worked with 速腾128线 before, does it have a similar reflection value as velodyne64? The model uses the reflection as part of the feature as well. You might set input_features in the config file as 0 and ret-train the model. Also, if you have some labels, you might want to use run.py with vis_level > 2 and check if at least your labels are correctly visible in the open3D.

GXQ0527 commented 3 years ago

用自己的点云数据去检测问题已经解决，kitti的bin文件保存的点云数据，其将反射强度归一化了，只要将自己的数据里面反射强度归一化即可。我还有一个问题想请教您，在代码里如何保存点云检测效果图，或者点云检测的鸟瞰图，open3d，具体操作在哪里？期待您的回复。

WeijingShi commented 3 years ago

Hi @GXQ0527, Glad to know you find the intensity normalization problem. Do you mind sharing a screenshot of the detection results? It would very interesting to see.

You can press the key P in open3d to save a screenshot (press`h to see other commands). If you want to code to save the image, the link maybe helpful.

GXQ0527 commented 3 years ago

ScreenCapture_2021-01-20-09-31-01 ScreenCapture_2021-01-20-09-32-18 这是结果图。

在我按下p之后，也有一个.json文件生成，里面内容如下，是相机的内参和外参吗？图片分辨率？ Screenshot from 2021-01-20 09-39-04

这些跟我输入的参数不一样呀？ Screenshot from 2021-01-20 09-43-18 期待您的回复！

WeijingShi commented 3 years ago

Hi @GXQ0527,

I think the json file is intrinsics/extrinsics for the open3D viewing windows.

The results look nice. Thank you for sharing. It is very helpful to others who also want to adapt their custom dataset. If I understand correctly, the key is to normalize the lidar intensity value to [0, 1] just like Kitti, correct?

Thanks,

GXQ0527 commented 3 years ago

是的，kitti点云数据集将反射强度进行了归一化。 作者，您好！我还有一个问题想请教您，在代码中如何输出像论文里面的点云检测图 1

或者这样的图 2

WeijingShi commented 3 years ago

Hi @GXQ0527, the visualization in the paper uses some custom scripts using open3D, we manually select the viewing angle and get the screenshot. But it may also be done by a script I think.

lrving-ww commented 2 years ago

ScreenCapture_2021-01-20-09-31-01 ScreenCapture_2021-01-20-09-32-18 这是结果图。

在我按下p之后，也有一个.json文件生成，里面内容如下，是相机的内参和外参吗？图片分辨率？ Screenshot from 2021-01-20 09-39-04

这些跟我输入的参数不一样呀？ Screenshot from 2021-01-20 09-43-18 期待您的回复！

你好 我想问一下 用自己的点云数据R0_rect这个矩阵怎么得到

bingo9062 commented 2 years ago

@GXQ0527 请问您的问题解决了吗？我也遇到了128线自己的数据集，识别不出，方便联系吗？非常感谢