Suggestion: Docker support and missing submodules (Tested on x86_64/RTX 4090)

[FredyHKU]

Hi there,

First of all, thank you for sharing this project.

While trying to set it up, I found the build process to be quite challenging, primarily due to some missing dependencies in the thirdparty directory and specific environment requirements. To address this, I have made some improvements to ensure a successful build.

What I have done:

Resolved Dependencies: Identified and added the missing submodules in the thirdparty folder required for a successful compilation.
CMake Adjustments: Made minor modifications to CMakeLists.txt to ensure compatibility with modern hardware and library paths.
Dockerization: Created a Docker environment specifically tested and verified on an x86_64 CPU platform with an NVIDIA RTX 4090 GPU.
Current Limitations:
While the project now compiles successfully in this environment, I have not been able to perform functional testing (such as running the Replica dataset) because the repository currently lacks detailed usage instructions or a testing guide.

Proposed Contribution:
If you are open to it, I would be happy to submit a Pull Request (PR) containing the Dockerfile, the necessary CMake adjustments, and the updated submodule configurations.

I hope this contribution can serve as a baseline for others, and perhaps the community or maintainers could help supplement the documentation and testing procedures for the datasets.

Looking forward to your thoughts!

Best regards,

YUAN Ye, Fred
City University of Hong Kong

The Docker files and image build scripts are attached in Gist below:
https://gist.github.com/FredyHKU/e8cb33c6ddb0113a3473cfb56c668c18

[FredyHKU]

DDN-SLAM Docker Compilation: Modifications Summary

Base Project: DDN-SLAM
Reference Project: Orbeez-SLAM

---

1. Third-Party Libraries

The original DDN-SLAM only defines git submodules but doesn't include actual code. For Docker build, we need to manually add these libraries:

Library	Version	Purpose
g2o	latest	Graph optimization framework
Sophus	1.22.4	Lie group algebra, must use C++14 compatible version for NVCC
Pangolin	v0.6	3D visualization, must use C++14 compatible version
instant-ngp-kf	-	Neural radiance field with keyframe support
DBoW2	-	Already included in original repo
Vocabulary	ORBvoc.txt	Needs decompression from .tar.gz

---

2. Files Supplemented from Orbeez-SLAM

The following example files are missing in DDN-SLAM but exist in Orbeez-SLAM:

Examples/RGB-D/rgbd_replica.cu
Examples/RGB-D/rgbd_scannet.cu
Examples/RGB-D/rgbd_tum.cu

Reason: DDN-SLAM repo only has partial example files. These are required to run on TUM/Replica/ScanNet datasets.

---

3. Code Modifications

3.1 CMakeLists.txt - Add CUDA Driver Library Linking

find_library(CUDA_DRIVER_LIB cuda
    PATHS /usr/local/cuda/targets/x86_64-linux/lib/stubs
          /usr/local/cuda/lib64/stubs
          /usr/lib/x86_64-linux-gnu
    REQUIRED)

target_link_libraries(rgbd_tum ORB_SLAM3 ngp tiny-cuda-nn ${CUDA_DRIVER_LIB})

Reason: tiny-cuda-nn uses CUDA Driver API for virtual memory allocation. In Docker environment, the CUDA driver library (libcuda.so) is not in the standard path—it's injected by NVIDIA Container Runtime at runtime. During compilation, we must link to the stub library (/usr/local/cuda/lib64/stubs/libcuda.so) to satisfy the linker, otherwise linking fails with "undefined reference to cuMemCreate" errors.

3.2 CMakeLists.txt - Disable Missing Example

# Commented out:
# add_executable(rgbd_inertial_openloris ...)

Reason: Source file Examples/RGB-D-Inertial/rgbd_inertial_openloris.cu doesn't exist in the repo.

3.3 OptimizableTypes.h / OptimizableTypes.cu - g2o Type Name Fix

// Before:
g2o::Vector7d

// After:
g2o::Vector7

Reason: Newer versions of g2o library renamed Vector7d to Vector7. This causes compilation errors like "Vector7d is not a member of g2o".

---

4. Docker Build Configurations

4.1 Eigen Version Replacement

rm -rf instant-ngp-kf/dependencies/eigen
ln -s /usr/local/include/eigen3/Eigen instant-ngp-kf/dependencies/eigen/Eigen

Reason: instant-ngp-kf bundles Eigen 3.4.90 which contains EIGEN_INTERNAL_HEADER_CHECK macro guards. These guards are incompatible with NVCC compiler, causing errors like "this file should not be included directly". System-installed Eigen 3.4.0 doesn't have this issue.

4.2 CUDA Library Symlinks

ln -s libcublasLt.so.11 libcublasLt.so.12
ln -s libcublas.so.11 libcublas.so.12

Reason: Host machine's NVIDIA driver (newer version) injects requests for CUDA 12 libraries, but Docker image uses CUDA 11.8. This mismatch causes "libcublasLt.so.12: cannot open shared object file" errors. Symlinks resolve this compatibility issue.

4.3 Sophus Build Flag

cmake .. -DSOPHUS_USE_BASIC_LOGGING=ON

Reason: Default Sophus uses fmt library for logging, which has template compatibility issues with NVCC. SOPHUS_USE_BASIC_LOGGING switches to simple std::cout logging, avoiding these compilation errors.

4.4 Build libyolov5.so

nvcc -shared -o libyolov5.so -Xcompiler -fPIC \
     preprocess.cu yololayer.cu \
     -DAPI_EXPORTS -lnvinfer -lcudart

Reason: yolo_ros project only builds an executable, not a shared library. DDN-SLAM needs to link against libyolov5.so for YOLO detection functionality.

---

5. Summary of Error → Fix Mapping

Compilation Error	Root Cause	Fix
undefined reference to cuMemCreate	CUDA driver API not linked	Add ${CUDA_DRIVER_LIB} to link libraries
Vector7d is not a member of g2o	g2o API changed	Change Vector7d → Vector7
EIGEN_INTERNAL_HEADER_CHECK error	Eigen 3.4.90 incompatible with NVCC	Replace with Eigen 3.4.0
libcublasLt.so.12 not found	CUDA version mismatch	Create symlinks .so.11 → .so.12
fmt template errors in Sophus	fmt library incompatible with NVCC	Use SOPHUS_USE_BASIC_LOGGING
libyolov5.so not found	Library not built	Manually compile from CUDA sources