PyTorch-implemented nerf-based models

Models

This repo provides PyTorch code for models that use the "NeRF" method to generate 3D images of objects. Currently, it supports 4 different models:

• NeRF
• NeuS
• TensoRF
• Nerfies

Hardware

The models are trained using following hardware:

• GPU: GTXX 1660 SUPER
• RAM: 32GB
• CPU: Ryzen 7 5700X

Result

• Performance

	NeRF	NeUS	Nerfies	TensoRF
Training speed	4it/s	5it/s	4it/s	51.6it/s
Rendering speed	21s/img	60s/img	-	-

NeRF

TensoRF

NeuS

Pipeline benchmark : NerfStudio

- Target: DataManager, Model pipeline is customized upon our environment - NOTE! RayGT and RayOutputs are currently dictionaries. In the future, they will be typed objects. - Component - RayBundles - the input to the forward pass of the Model. needed for both training and inference time - describe origin, viewing directions, which are used in rendering process - RayGT objects, however, are needed only during training to calculate the losses in the Loss Dict. - information like pixel ground truths, whic are used in loss computation

---

References

1. pytorch repo : link
2. volumetric rendering in NeRF link
3. Camera calibration link
4. NeUS link
5. NeRF Studio
   • blog, Docs
   • Camera model
     • Instrinsic/Extrinsic
     • Equirectangular/Spherical Camera Model
     • Distortion Parameters
   • Sample representation
   • Ray samplers
     • Strafied Sampling, Heirarchical Sampling
   • Spatial distortions
   • • Encoders