Segmentation fault (Address not mapped to object [0x70]) When "rosrun loop_fusion loop_fusion_node"

[chenjack9017]

Your paper is excellent! But now I am facing some difficult problems to deal with

1、Roscore
2 roslaunch vins vins_rviz.launch
3 rosrun vins vins_node /home/cjw/ws1/src/D_VINS-main/config/euroc/euroc_mono_imu_config.yaml 4 rosrun loop_fusion loop_fusion_node /home/cjw/ws1/src/D_VINS-main/config/euroc/euroc_mono_imu_config.yaml
5 rosbag play ~/SLAM/MH_04_difficult.bag

After returning the rosbag, the fourth window runs normally, but waiting for a while will result in the following error:
(base) cjw@c:~/ws1$ rosrun loop_fusion loop_fusion_node /home/cjw/ws1/src/D_VINS-main/config/euroc/euroc_mono_imu_config.yaml config_file: /home/cjw/ws1/src/D_VINS-main/config/euroc/euroc_mono_imu_config.yaml vocabulary_file/home/cjw/ws1/src/D_VINS-main/loop_fusion/../support_files/brief_k10L6.bin loop start load vocabulary BRIEF_PATTERN_FILE/home/cjw/ws1/src/D_VINS-main/loop_fusion/../support_files/brief_pattern.yml cam calib path: /home/cjw/ws1/src/D_VINS-main/config/euroc/cam0_mei.yaml VIO input, perfrom 4 DoF (x, y, z, yaw) pose graph optimization no previous pose graph Stack trace (most recent call last) in thread 113307: #6 Object "[0xffffffffffffffff]", at 0xffffffffffffffff, in #5 Object "/lib/x86_64-linux-gnu/libc.so.6", at 0x7f19c79f9352, in clone #4 Object "/lib/x86_64-linux-gnu/libpthread.so.0", at 0x7f19cd07c608, in #3 Object "/lib/x86_64-linux-gnu/libstdc++.so.6", at 0x7f19c7d0edf3, in #2 Object "/home/cjw/ws1/devel/lib/loop_fusion/loop_fusion_node", at 0x55625289f97f, in process() #1 Object "/home/cjw/ws1/devel/lib/loop_fusion/loop_fusion_node", at 0x5562528deade, in KeyFrame::KeyFrame(double, int, Eigen::Matrix<double, 3, 1, 0, 3, 1>&, Eigen::Matrix<double, 3, 3, 0, 3, 3>&, cv::Mat&, std::vector<cv::Point3_<float>, std::allocator<cv::Point3_<float> > >&, std::vector<cv::Point_<float>, std::allocator<cv::Point_<float> > >&, std::vector<cv::Point_<float>, std::allocator<cv::Point_<float> > >&, std::vector<double, std::allocator<double> >&, int) #0 Object "/home/cjw/ws1/devel/lib/loop_fusion/loop_fusion_node", at 0x5562528dc63a, in KeyFrame::computeWindowSuperpoint() Segmentation fault (Address not mapped to object [0x70]) 段错误 (核心已转储)
Have you encountered a similar problem before?

[kajo-kurisu]

use euroc_stereo_imu config instead or follow it to revise your config

[chenjack9017]

Thank you very much for your reply！
The environment I am currently using is Ubuntu 20.04 and Ros1, and all other environments have been designed according to your requirements. However, the issue you mentioned in issue 3 has arisen
[2024-10-21 17:30:10][fatal][trt_tensor.cpp:359]:Assert failed, ndims == shape_.size()
Then I followed your instructions to view the dimensions and printed the values for each dimension:
`cout << "Binding Name: " << bindingName << endl;

        cout << "Dimensions: ";
        for (int j = 0; j < dims.nbDims; ++j) {
        cout << dims.d[j] << " "; 
         }
        cout << endl;`

The output is:
cjw@c:~$ rosrun loop_fusion loop_fusion_node /home/cjw/dvins/src/D_VINS/config/euroc/euroc_mono_imu_config.yaml
Binding Name: img
Binding Name: des
Binding Name: kpts0
Binding Name: kpts1
Binding Name: desc0
Binding Name: desc1
Binding Name: mscores0
Binding Name: mscores1
[2024-10-21 17:30:10][fatal][trt_tensor.cpp:359]:Assert failed, ndims == shape_.size()

I am confident that the config was downloaded according to your file and the v0.1.3 version of superpoint_512 and superpoint_lightglue as per your request.
I'm really not sure which direction to look for the mistake from now. Can you give me some guidance?

[chenjack9017]

I'm very sorry, the code in trt_inner is:
` cout << "Binding Name: " << bindingName << endl;

        cout << "Dimensions: ";
        for (int j = 0; j < dims.nbDims; ++j) {
        cout << dims.d[j] << " "; 
         }
        cout << endl;`

[chenjack9017]

and this is
cjw@c:~$ rosrun loop_fusion loop_fusion_node /home/cjw/dvins/src/D_VINS/config/euroc/euroc_stereo_imu_config.yaml Binding Name: img Dimensions: 1 3 320 320 Binding Name: des Dimensions: 1 512 Binding Name: kpts0 Dimensions: 1 -1 2 Binding Name: kpts1 Dimensions: 1 -1 2 Binding Name: desc0 Dimensions: 1 -1 256 Binding Name: desc1 Dimensions: 1 -1 256 Binding Name: mscores0 Dimensions: 1 -1 [2024-10-20 23:29:25][fatal][trt_tensor.cpp:359]:Assert failed, ndims == shape_.size()

[kajo-kurisu]

Make sure you really follow the steps I asked you to do, and check that your model was imported correctly at the beginning of deep_net.h

[chenjack9017]

Thank! I will try it again! kajo-kurisu ***@***.***>于2024年10月22日 周二09:24写道：

…

Make sure you really follow the steps I asked you to do, and check that your model was imported correctly at the beginning of deep_net.h — Reply to this email directly, view it on GitHub <#11 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BHWEYHBV4Z4NSSXHWZM3UILZ4WSMNAVCNFSM6AAAAABQBUH3J2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIMRYGAZDGNZYGA> . You are receiving this because you modified the open/close state.Message ID: ***@***.***>

[chenjack9017]

Make sure you really follow the steps I asked you to do, and check that your model was imported correctly at the beginning of deep_net.h

师兄您好，我现在貌似查到了大概是什么地方不对，但是不太确定。
您在trt_infer.cpp文件中要求if( !strcmp(bindingName,"matches0") || !strcmp(bindingName,"mscores0") || !strcmp(bindingName,"scores_r") )他们这几个输出的维度的第0维为动态的，但是我在检查了您在issue3中提出直接下载他们v0.1.3版本中的那两个onnx文件，其中的superpoint_lightglue.onnx转换成engine之后，我发现他们的第1维就不是您所要求的-1了，而是变成了1，其他的文件的数据貌似都是符合您在文件中所说的要求；

师兄您知道怎么解决吗？万分感谢！

[kajo-kurisu]

onnx的维度示意，检查一下你是不是自己导出错误了。。。

[kajo-kurisu]

而且你的问题是出自 computewindowsuperpoint，多半是你的sp_re没弄对

[chenjack9017]

onnx的维度示意，检查一下你是不是自己导出错误了。。。

师兄，非常抱歉，这个onnx是直接下载的他们v0.1.3版本的superpoint_lightglue.onnx吗？我下载的跟这个不一样呜呜呜

[chenjack9017]

而且你的问题是出自 computewindowsuperpoint，多半是你的sp_re没弄对

师兄，这个是sp_re的onnx信息

[chenjack9017]

@kajo-kurisu 感谢师兄，我现在已经跑通了euroc_stereo_imu_config.yaml的工作，显示的非常好

但是我在运行单目的时候还是会显示错误（虽然我可以不用单目，但是就是想求知一下，感谢您）

sp_re也是在pytorch1.13+cu11.7的环境生成的onnx，外加和您要求的转换引擎指令相同。

ps：您之前说的关于superpoint_lightglue.onnx的文件，从v0.1.3版本的release直接下载的和您要求的输出不一致，但是我后面无意发现他们在v2.0当中所使用的输出代码就只有那两个，因此我是在pytorch2.1+cu121的环境下，使用他们的代码进行了onnx的导出。随后再次使用cu11.7进行的转换engine。最后的的结果就是双目imu运行的第一个图所示，但是单目如图2所示依旧是失败了。

不知道有没有哪里不太对的地方，如果有的话非常感谢您的帮助，让我受益匪浅！
如果没有的话更感谢您对我的回复与指导，您的工作非常好！祝您学业进步！再次感谢！

[kajo-kurisu]

单双目按理说，只要你的config跟双目中的配置格式一样的话，是不影响模型工作的，这里报错是因为sp_re推理上出了问题，你可以检查一下单双目下的sp_re函数的输入区别，还有换数据集的时候width_adj ,height_adj要相应的改变。最后，vins下有euroc_mono 的 config，看看你是否用错了

[kajo-kurisu]

我使用的是github 上一个 kaist2bag 的转换工具，只使用了图像和imu数据，运行时订阅的是imu_raw 的数据。
关于误差比较大的问题，之前做测试的时候大致检查过，可能是由于过度优化，导致一些车行驶的时候本来就是弯曲的路线给优化直了，最终的误差较大，但是基本上多次运行的平均误差与论文内展示的数据一致。另外由于运行环境的不同，你可以试试调节PNP的参数和回环阈值来选出比较适合你的参数