16889 Assignment 3: Volume Rendering and Neural Radiance Fields.
Xikai Dai
Differentiable voluem rendering
Pi: Vis_Grid & Vis_Ray
1.4: Point Sampling
Pic: Point visulization
1.5: Volume Rendering
Pit: Part 1 rendering and Depth Visualization
2: Optimization, loss and training
The length and center value are:
Box center: (0.2497246563434, 0.250206083059, -0.000905538094229)
Box side lengths: (2.003785848617, 1.500813484191, 1.50295567512)
With the Gif 
3: Optimizing a Neural Radiance Field
Pic: NeRF without the view dependence
4: NeRF Extra
4.1 with View Dependence
Pic: NeRF with the view dependence
The addtion of view depence as input will increase the prediction accuracy under the same number of epoch training. But, based on the loss value during the training, it makes the overall network quicker to decrese the loss, and further cause the overfitting for longer training. Most of the network structure is set to be the same between Part3 and Part4.1. In terms of the image generation quality, the view dependence is slightly better than the sample_points only training, but due to the low resolution of the generated image, the difference is hard to observe.
4.3 High resolution
Pic: part3_low_resolution comapre to 4.3 high resoltuion comparison
pic: high resolution gif 5fps
The High resolution network is different than part.3 mostly due to the training time of the process. One of the factor is the vislauzaiton of gif in between of the training, To lower the training time, the number of camera_views are decreased (from 20 to 5). But, based on the screenshot of the gif at the same zoom resolution, the Hight resolution is significantly better comapre to the low res one in part.3 During the training, the n_pts_per_ray will effect the render result, and it is set to be 200 rather than default 128.
