iNeRF
Motivation
why invert???
How to obtain an initial pose estimation?
given an image, find the translation and rotation of a camera relative to a 3D object or scene.
Setting
no mesh object model
one image
using gradient descent to minimize the residual
Experiment
how different batch sizes of rays affect iNeRF on a synthetic dataset
We then show that for complex real-world scenes from the LLFF dataset, iNeRF can improve NeRF by estimating the camera poses of novel images and using these images as additional training data for NeRF
how to sample rays during pose refinement for iNeRF to collect informative gradients
Finally, we show iNeRF can perform category-level object pose estimation, including object instances not seen during training, with RGB images by inverting a NeRF model inferred from a single view
INeRF formulation
- Gradient-Based SE(3) Optimization
3.Self-supervised NeRF with iNeRF
NeRF: T, I —> theta
iNeRF: I, theta —> T
click to edit
click to edit
click to edit
2)use iNeRF to take in additional unknown-pose observed images and solve for estimated poses
1)train a NeRF given a set of training
RGB images with known camera poses
click to edit
- Sampling Rays
click to edit
3) add 2) to 1) to do a semi-supervised learning
click to edit
click to edit
click to edit
click to edit
click to edit
click to edit
click to edit
click to edit
click to edit
click to edit
click to edit
click to edit
click to edit
Experiments
LLFF Dataset
Sim2Real Cars
Synthetic dataset
For each scene, choose 5 test images and generate 5 different pose initializations???
click to edit
results: report pose error with less than 5 degrees or 5cm
click to edit
click to edit
ShapeNet-SRN Cars
click to edit
click to edit
Dataset Details:
1) using the "car" classes of ShapeNet introduced by Stizmann including 3514 cars.
2) split the 3514 cars into training, validation and test set.
3) for objects in the test set, we render an image in an archimedean spiral and then select another image within certain offset.