NIEO
A neural implicit representation method for LiDAR mapping that defines an energy functional on the shape operator of a signed distance field as a regularization term, encoding surface smoothness priors into network weights during training with no additional overhead at inference.
Timeline (last 90 days)
The key innovation of the NIEO method is to regularize the SDF with a shape operator energy functional defined on the SDF level set.
The energy functional regularization of NIEO is performed only during training; at inference time only a standard forward pass is done, introducing no additional inference time overhead.
Unlike point cloud post-processing filtering, NIEO encodes smoothness into the network weights, so no curvature, shape operator, or energy term needs to be computed at inference time.
Experiments on large-scale datasets show that the NIEO method can significantly improve the smoothness of reconstructed surfaces while preserving geometric details.
The NIEO code has been made public on GitHub, with the repository named NIEO (Neural Implicit Energy Optimization) at https://github.com/HrishikeshVish/nieo.
NIEO does not introduce a more complex network structure; instead, it applies the idea of energy functionals from classical geometric optimization to the neural parameter space.
All Facts (6)
The key innovation of the NIEO method is to regularize the SDF with a shape operator energy functional defined on the SDF level set.
50%UnverifiedThe energy functional regularization of NIEO is performed only during training; at inference time only a standard forward pass is done, introducing no additional inference time overhead.
50%UnverifiedUnlike point cloud post-processing filtering, NIEO encodes smoothness into the network weights, so no curvature, shape operator, or energy term needs to be computed at inference time.
50%UnverifiedExperiments on large-scale datasets show that the NIEO method can significantly improve the smoothness of reconstructed surfaces while preserving geometric details.
50%UnverifiedThe NIEO code has been made public on GitHub, with the repository named NIEO (Neural Implicit Energy Optimization) at https://github.com/HrishikeshVish/nieo.
50%UnverifiedNIEO does not introduce a more complex network structure; instead, it applies the idea of energy functionals from classical geometric optimization to the neural parameter space.
50%