Please enable JavaScript.
Coggle requires JavaScript to display documents.
iMAP: Implicit Mapping and Positioning in Real-Time - Coggle Diagram
iMAP: Implicit Mapping and Positioning in Real-Time
3
rd
: Depth and colour rendering
5
th
: Keyframe selection
1.The first frame: initialize the network and fix the world
coordinate frame.
lock a copy of the network and present a snapshot of 3D
map.
new region checked against the copy.
The strategy for keyframe selection.
4
th
: Joint optimization
1
st
: system overview
2
nd
:implicit scene neural network
(x,y,z) -->
4 hidden layers (feature size 256) + 2 output
--> color + volume density
Active Sampling
6
th
: Active Sampling
Keyframe Active Sampling
1) distribution for sequence
2) distribution for pixels in keyframe
Image Active Sampling
1) 200 per image
2) loss distribution is calculated from a set of uniform samples
3) active samples are further allocated proportional to the loss distribution
Bounded Keyframe Selection
1) 3 keyframes at each iteration
2) W = 5
contributions
2)
The ability to incrementally train an implicit scene network in real-time, enabled by automated keyframe selection and loss guided sparse active sampling
3)
A parallel implementation (fully in PyTorch with multi-processing) of our presented SLAM formulation which works online with a hand-held RGB-D camera
1)
The first dense real-time SLAM system that uses an implicit neural scene representation and is capable of jointly optimising a full 3D map and camera poses.