1. pose estimation used in PAD, nips23
2. use mse loss of sampled pixels;

render-and-compare strategy

1. a photometric loss on sample pixels.
   1. 这具体定了本方法，只能做初值好的情形。
   2. why sampling? expensive computational cost of NeRF rendering
      1. b = 2048 rays per gradient step, which corresponds to a single forward/backward pass that fits within GPU memory and provides 150× faster gradient steps on a 640 × 480
         1. takes around 20 seconds to run 100 optimization steps,
   3. Fig. 3: An illustration of 3 sampling strategies.

Related work

Object Pose Estimation from RGB Images

require access to objects’ 3D models during both training and testing

Limitations

lighting and occlusion can severely affect the performance of iNeRF

References

iNeRF: Inverting Neural Radiance Fields for Pose Estimation, IROS21

iComMa, 24