DynamicFusion: Reconstruction and tracking of non-rigid scenes in real-time
read more
Citations
Past, Present, and Future of Simultaneous Localization and Mapping: Toward the Robust-Perception Age
Past, Present, and Future of Simultaneous Localization And Mapping: Towards the Robust-Perception Age
VNect: real-time 3D human pose estimation with a single RGB camera
Deferred neural rendering: image synthesis using neural textures
VNect: Real-time 3D Human Pose Estimation with a Single RGB Camera
References
KinectFusion: Real-time dense surface mapping and tracking
A volumetric method for building complex models from range images
KinectFusion: real-time 3D reconstruction and interaction using a moving depth camera
Efficient Model-based 3D Tracking of Hand Articulations using Kinect
Real-time 3D reconstruction at scale using voxel hashing
Related Papers (5)
Frequently Asked Questions (19)
Q2. What is the key extension over static state space representations?
The warp-field is the key extension over static state space representations used in traditional reconstruction and SLAM systems, and its estimation is the enabler of both non-rigid tracking and scene reconstruction.
Q3. What is the main computational complexity in minimisingE?
The main computational complexity in minimisingE involves constructing and factorizing the Gauss-Newton approximation of the Hessian: J>J = J>dJd + λJ > r Jr. First, the authors note that the k-nearest node field induces non-zero blocks in the data term component J>dJd for each pair of nodes currently involved in deforming V into an observable region of the live frame.
Q4. What is the main focus of non-rigid tracking research?
The vast majority of non-rigid tracking research focuses on human body parts, for which specialised shape and motion templates are learnt or manually designed.
Q5. What is the main reason why the authors are building the full linear system on the GPU?
Building the full linear system on the GPU currently hinders real-time performance due to requirements on global GPU memory read and writes.
Q6. How do the authors compute the next level of regularisation nodes?
The authors compute the next l = 1 level of regularisation nodes by running the radius search based sub-sampling on the warp field nodes dgv to an increased decimation radius of βl, where β > 1, and again compute the initial node transforms through DQB with the now updated Wt.
Q7. What is the projective signed distance at the warped canonical point?
The projective signed distance at the warped canonical point is:psdf(xc) = [ K−1Dt (uc) [ u>c , 1 ]>] z − [xt]z , (3)where uc = π (Kxt) is the pixel into which the voxel center projects.
Q8. How do the authors obtain an initial estimate for data-association between the model geometry and the live frame?
The authors obtain an initial estimate for data-association (correspondence) between the model geometry and the live frame by rendering the warped surface V̂w into the live frame shaded with canonical frame vertex positions using a rasterizing rendering pipeline.
Q9. What is the way to reconstruct a fixed topology?
An intriguing approach to template-free non-rigid alignment, introduced in [17] and [26], treats each nonrigid scan as a view from a 4D geometric observation and performs 4D shape reconstruction. [30, 29] reconstruct a fixed topology geometry by performing pair-wise scan alignment. [24] use a space-time solid incompressible flow prior that results in water tight reconstructions and is effective against noisy input point-cloud data. [28] introduce animation cartography that also estimates shape and a per frame deformation by developing a dense correspondence matching scheme that is seeded with sparse landmark matches.
Q10. How can the authors avoid the weight of the deformation node?
since the associated weight of each deformation node reduces to a very small value outside of 3dgw, any data term there can be safely ignored.
Q11. What is the k-nearest transformation node to the point x?
N (x) are the k-nearest transformation nodes to the point x and wk : R3 7→ R defines a weight that alters the radius of influence of each node and SE3(.) converts from quaternions back to an SE(3) transformation matrix.
Q12. What are the core algorithmic components to the system that are performed in sequence on arrival of each?
There are three core algorithmic components to the system that are performed in sequence on arrival of each new depth frame:1. Estimation of the volumetric model-to-frame warp field parameters (Section 3.3)2.
Q13. What are the two categories of non-rigid tracking?
While no prior work achieves real-time, template-free, non-rigid reconstruction, there are two categories of closely related work: 1) real-time non-rigid tracking algorithms, and 2) offline dynamic reconstruction techniques.
Q14. What is the way to track a low resolution shape?
Other techniques directly track and deform more general mesh models. [12] demonstrated the ability to track a statically acquired low resolution shape template and upgrade its appearance with high frequency geometric details not present in the original model.
Q15. What is the TSDF for a point in the canonical frame?
This allows the TSDF for a point in the canonical frame to be updated by computing the projective TSDF in the deforming frame without having to resample a warped TSDF in the live frame.
Q16. What is the difference between the two warp functions?
The authors note that scaling of space can also be represented with this warp function, since compression and expansion of space are represented by neighbouring points moving in converging and diverging directions.
Q17. What is the process of updating the deformation graph nodes?
This consists of incrementally updating the deformation graph nodes Nwarp, and then recomputing a new hierarchical edge topology E that expands the regularisation to include the new nodes.
Q18. What is the definition of the knearest neighbours of each node?
In original applications of the embedded deformation graph [25] approach to non-rigid tracking, E is defined as the k−nearest neighbours of each node or all nodes within a specified radius.
Q19. What is the optimisation required to estimate the canonical surface geometry?
In the preceding subsections the authors defined how the canonical space can be deformed through W (Section 3.1), introduced the optimisation required to estimate warp-field state through time (3.3), and showed how, given an estimated warp field, the authors can incrementally update the canonical surface geometry (3.2).