Once we have a set of cameras, we want to cast camera rays associated with each pixel.
Along these ray we will sample the field and aggregate the samples to predict the pixels value (ie. color). The parameterization of the samples are described here however we must decide where to place these samples along a ray. For this task we will use a
In the ideal world we would compute many dense samples along a ray. Unfortunatly each additional sample adds a computation cost to the system as it needs to be processed by the field which is often a neural network.
As a result it is common for NeRF methods to use on the order of 100 samples. Therefore we want to optimize where those samples are placed in the scene.
For example, if the sene can be bounded by a box and the objects are all similar scales, uniform sampling along the ray may be a good option. On the other hand if the scene is unbounded (potential extending as far as the the eye can see) uniform sampling does not make sense as the samples would be very sparse for close objects. In this case a different sampling like Uniform in Disparity may perform better.
Most samplers has the option to stratify the samplers. When stratified, each sample is randomly perturbed.
The magnitude of the pertubation is such that the sample ordering remains consistent and the overal distribution statistics are not changed. Using stratified samples during training generally improves the reconstructions as it help prevent overfitting.
During inference stratified sampling should be disabled (nerfstudio samplers will do this) as it can cause noisy artifacts when the camera moves.
It is important to sample the scene where it has content otherwise the reconstruction quality will be reduced.
One trick that is often employed in NeRF methods is to do multiple round of sampling. The first round can use a predefined sampler (ie. Uniform) to generate an image. Once the space is sampled, we have an idea which samples contributed to the final color.
We can use this information to sample more around those regions using a
PDFSampler. The PDF sampler is described more below.
These are the most basic samplers that spaces samples based on a predefined function. These samplers have all have a starting and ending distance (also known as a near/far plane). The plots below are histograms of points sampled from some predefined samplers.
# COLLAPSED import torch from nerfstudio.cameras.rays import RayBundle from nerfstudio.model_components import ray_sampler import plotly.graph_objects as go from plotly.subplots import make_subplots num_samples = 1000 near = 2 far = 5 train_stratified = False samplers = [ ray_sampler.UniformSampler, ray_sampler.LinearDisparitySampler, ray_sampler.SqrtSampler, ray_sampler.LogSampler, ] fig = make_subplots( rows=2, cols=2, subplot_titles=("Uniform", "Linear in Disparity", "Square Root", "Log Sampler"), shared_xaxes=True, shared_yaxes=True, vertical_spacing=0.1, ) for i, Sampler in enumerate(samplers): sampler = Sampler(num_samples=num_samples, train_stratified=train_stratified) ray_bundle = RayBundle( origins=torch.ones([1, 3]), directions=torch.ones([1, 3]), pixel_area=torch.ones([1, 1]), nears=torch.ones([1, 1]) * near, fars=torch.ones([1, 1]) * far, ) samples = sampler.generate_ray_samples(ray_bundle) trace = go.Histogram(x=samples.frustums.starts[0, :, 0], nbinsx=50) fig.append_trace(trace, i // 2 + 1, i % 2 + 1) fig.update_yaxes(title_text="# Samples", row=1, col=1) fig.update_yaxes(title_text="# Samples", row=2, col=1) fig.update_xaxes(title_text="Distance", row=2, col=1) fig.update_xaxes(title_text="Distance", row=2, col=2) # Overlay both histograms fig.update_layout(height=700, hovermode=False, showlegend=False, margin=dict(l=20, r=20, t=50, b=20)) fig.update_yaxes(range=[0, 80]) fig.update_traces(opacity=0.7) fig.show()