Directional Sampling Overview

(starting at 4:56:25)

Transcript

1. Title Slide

Hello, my name is Stephen Wasilewski and I am presenting some work I have prepared along with my co-authors. Raytraverse is a new method that guides the sampling process of a daylight simulation.

2. The Daylight Simulation Process

To understand how this method can enhance the daylight simulation process, it is useful to view the process by parts.

2.b

The model describes how geometry, materials, and light sources are represented.

2.c

Sampling determines how the analysis dimensions are subdivided into discrete points to simulate.

2.d

These views rays are solved for by a renderer, yielding a radiance or an irradiance value for each view ray.

2.e

This output is evaluated according to some metric or otherwise preparing the data for interpretation.

3. Assumptions

To make a viable workflow, each of these parts require (whether explicitly or implicitly) a number of assumptions that define the limitations and opportunities of the method. To explain this in practical terms, here are three examples of well known climate based modeling methods for visual comfort.

4. CBDM Methods for Visual Comfort: Ev based

Illuminance based methods, including DGPs (simplified Daylight Glare Probability), limit the directional sampling resolution to a single sample per view direction in order to efficiently sample a larger number of positions and sky conditions throughout a space.

Unfortunately: Even if the employed rendering method perfectly captures the true Illuminance, as a model for discomfort glare it fails to account for scenes where the dominant driver of discomfort is contrast based or due to small bright sources in an otherwise dim scene.

5. CBDM Methods for Visual Comfort: 3/5 Phase

The 3-phase and 5-phase methods focus on the model and render steps. These methods fix the implementations of the material and sky models by discretizing the transmitting materials and sky dome in order to replace some steps of the rendering process with a matrix multiplication.

6. CBDM Methods for Visual Comfort: eDGPs

Like the 5-phase method, The enhanced-simplified daylight glare probability method, developed to overcome the limitations of illuminance only metrics, uses separate sampling and rendering assumptions for the indirect contribution and direct view rays. The adaptation level is captured by an illuminance value, but glare sources are identified with an image calculated for direct view ray contributions only.

7. Existing Options For Sampling a Point

In all of these methods, the sampling is treated as a fixed assumption.

7.b

Either directional sampling is directly integrated into an illuminance by the renderer,

7.c

or a high resolution image is generated.

7.d

This is because at intermediate image resolutions the accuracy of the results can be worse than an illuminance sample, and are unreliable for capturing contrast effects due to small sources.

7.e

So unlike sampling positions or timesteps which can be set at arbitrary spacing and easily tuned to the needs of the analysis, directional sampling is much more of an all or nothing choice; where the additional insights offered by an image can require 1 million times more data than a point sample. But is this really necessary?

7.f

Whether through direct image interpretation or any of the commonly used glare metrics, the critical information embedded in an HDR image is usually simplified to a small set of sources and background, each with a size, direction and intensity. We cannot directly sample this small set of rays because we do not know these important directions ahead of time, but how close can we get?

7.g

The raytraverse method provides a means to bridge the gap between point samples and high resolution images, allowing for a tunable tradeoff between simulation time and accuracy.

Our approach is structured by a wavelet space representation of the directional sampling. It works by applying a set of filters to an image to locate these important details.

8. Wavelet Decomposition

To match our sampling space, we apply these filters to a square image space based on the Shirley-Chiu disk to square transform, which preserves adjacency and area, both necessary for locating true details.

8.b

For each level of the decomposition, The high pass filters, applied across each axis (vertical, horizontal, and in combination) isolate the detail in the image, and the low pass filter performs an averaging yielding an image of half the size. This process is repeated, applying the high pass filters to the approximation, down to some base resolution. Each level of the decomposition stores the relative change in intensity at a particular resolution (or frequency).

8.c

The total size of the output arrays is the same as the original, and can be used to perfectly recover the original signal through the inverse transform.

The benefit to compression comes from the fact that the magnitude of the detail coefficients effectively rank the data in terms of their contribution to the reconstruction. By thresholding the coefficients, less important data can be discarded.

8.d

Even after discarding over 99% of the wavelet coefficients, the main image details are recoverable and only some minor artifacts have been introduced.

This property, that the wavelet coefficients rank the importance of samples at given resolutions, makes detail coefficients useful for guiding the sampling of view rays from a point.

9. Reconstruction Through Sampling

This process works as follows:

Beginning with a low resolution initial sampling the large scale features of the scene are captured.

Mimicking the wavelet transform, We apply a set of filters to this estimate and then use the resulting detail coefficients both to find an appropriate number of samples, and as probability distribution for the direction of these samples.

The new sample results returned by the renderer are used to update the estimate, which is lifted to a higher resolution.

This process is repeated up to a maximum resolution, equivalent to (or higher than) what a full resolution image might be rendered at.

10. Component Sampling

There are some cases where the wavelet based sampling will not find important details, such as specular views and reflections of the direct sun. Fortunately, because our method uses sky-patch coefficients to efficiently capture arbitrary sky conditions (similar to 3 phase and others), we can structure the simulation process in such a way to compensate for these misses. I refer you to our paper for details on how this works.

11. Results

Instead, I’ll spend my remaining time sharing a few examples of scenes captured with: our approach, a high resolution reference and a matching uniform resolution image to demonstrate the benefits of variable sampling.

In addition to image reconstructions, the relative deviation from the reference is shown for vertical illuminance (characterizing energy conservation) and UGR (Unified Glare Rating, characterizing contrast), relative errors greater than 10% are highlighted in red.

This very glary scene highlights the different paths that light takes from the sun to the eye, including direct views, rough specular and diffuse reflections of the sun and sky. While the deviation in the low resolution image is unlikely to change a prediction in this case, the large errors show a failure case for uniform low-res sampling.

11.b

A more complex, but also more likely scenario is that roller shades will be closed. While there are open questions on how to evaluate the specular transmission of such materials, raytraverse does not introduce any substantial new errors to this process.

11.c

Raytraverse performs similarly well for partially open venetian blinds.

11.d Including deeper in a space where the floor reflection dominates.

11.e

Raytraverse, without virtual sources or other rendering tricks, handles the case of specular reflections of the direct sun, a difficult problem for low resolution sampling.

11.f

One case that we would expect raytraverse to struggle with would be a high frequency pattern like the dot frit shown here. And while the sampling does miss parts of the pattern, especially the lower contrast areas, enough of the detail is caught to meaningfully understand the image and, because of the direct sun view sampling, maintains high accuracy.

11.g

In cases where more image fidelity is desired, raytraverse can be tuned to increase the sampling rate with a proportional increase in simulation time, but in our paper we show that the low sampling rates previously shown achieve a high level of accuracy for field of view metrics.

12. Thank you

Thank you for watching my presentation.