dkotfis / octree-slam Goto Github PK

Keep track of bounding boxes accessed to determine what parts of the tree are least recently updated.

Create a CUDA kernel that determines whether two points are valid correspondences based on their vertex, normal, and color distances.

Current control cannot translate the origin effectively and takes multiple clicks and drags to accomplish large rotations.
Suggest using more standard WASD + mouse control commonly found in first-person video games.

Alpha channel is currently 7 bits due to use of signed ints.
Would prefer switching to unsigned, and using 8 bits of alpha.

The brick pool is not supported in the new SVO. Bring it back. It will be useful for ray/cone tracing.

Voxelization grid size is currently determined at compile time.
Support dynamic checking to determine bounds of scene, as well as changing size when the scene becomes larger.

Use Least Recently Used (LRU) mechanism from GigaVoxels to copy unused octree data from GPU to CPU.

Hi Dave,

Great implementation! I would just like to confirm something.

After you have generated the normal map from a single depth frame, how is the normal map fused into the octree. Is it done by simply applying the current rotation matrix to each normal in the map and then setting the corresponding voxel's normal?

Thanks

Determine whether the GPU octree exceeds a threshold and pull to CPU the portion that is LRU.

Currently SVO uses voxelization parameters to determine grid sizes, etc. It should be a standalone utility that can construct an octree from arbitrary update types.

Currently duplicate a cube to construct a new VBO using CUDA.
Would be more efficient to skip this and use OpenGL instanced drawing to just render the same cube vbo at each center with each color.

Create a CUDA rendering method that casts rays into an SVO to render.

The current SVO structure only uses the node pool from GigaVoxels. Use texture/surface memory to add a brick pool capable of hardware interpolation for ray/cone tracing.

Compute freespace keys and update off of them (not just end-points).

Incorporate photometric ICP for visual odometry estimate.

Detect where a node pool or brick pool exceeds its allowed size. At least elegantly error and exit the program rather than crashing. Ideally, create a new structure and copy the old data into it.

Build all external source files into static libs, and replace them in the repo with these libs. Should not need to be compiling them as part of the project.

Add OpenNI2 dependencies to the project, and create an interface that receives RGB-D images and copies them to GPU memory for use in CUDA.

Implement octree ray-tracing for rendering without voxel extraction.

Currently only colors green. Add support for texture mapping for meshes where texture coordinates are available.

Add API function to SVO to allow camera rays to be used to update according to a sensor model. Use the alpha channel of the octree to store the net # hits, which can be used to reproduce log odds from only 8 bits.

The function needs to be device so it can be called by a kernel that parallelizes over the rays of an RGB-D image.

Compute the distance between points in a point cloud to determine the resolution at which the octree is updated.

Perform timing analysis on octree update process, and optimize things such as block/grid sizes to get better performance.

Perform iterative closest point calculation (in parallel with CUDA) using point correspondences. Perform Cholesky decomposition calculation on the CPU. Use the same minimization function as KinectFusion.

Create a CUDA kernel that uses an intrinsic camera calibration matrix to calculate a vertex map, then creates a normal map from adjacent vertices.