We have long admired the work of VFX artist Keith Adams, coming as he does from a very strong technical background, but like many outstanding compositors, he is also a fine art painter and produces stunning oil paintings.
We recently spoke in depth to Keith about his approach to shot design and it grew to be this exclusive look at the general issue and history of perspective work, matching it, designing for it and, for example, creating correct foreground and background plates for VFX work.
Keith Adams: Spin VFX Atlanta, founder and VFX supervisor
In an industry characterized by ever-advancing technologies, and limitless computing power, I find myself looking to old-school methods to solve visual effects problems. By old school, I am referring to 15th century, early-renaissance artisans and mathematicians who developed the theories of visual perspective. The ideas they pioneered go well beyond the elementary perspective theories taught in grade school. And they provide an objective means by which to make optically correct calculations allowing multiple elements in the same dimensional space. These principles can also be reverse-engineered applying the techniques to existing footage and still photos. Important variables such as camera placement, focal length, shadow location and relative sizing can be determined within allowable tolerances. These techniques bring the film professional outside the realm of guessing using the partial-mix and sharpie doodles on a monitor and moves us to a place where calculations and informed approximations can be made. Many aspects of a shot can be manipulated in post; however, correcting the perspective to match another is much more difficult and often not possible without undesirable results.
A brief history of visual perspective
Artist such as Robert Campin (1375-1444) and Paolo Uccello (1397 – 1475) as well as progressive thinkers such as Filippo Brunelleschi (1377 – 1446) are generally considered among the first to exploit the newfound theories. In their day, advancing ideas of visual perspective was akin to today’s cutting edge technologies in film and TV because it brought realism to their art. Forever, the way we see the world as represented in non-reality forms was improved. Up until the renaissance, artists generally painted distant objects smaller and near objects large. However, placement, specific size, shapes, angles and foreshortening were subjective and had little bearing on reality. Objects and environment had no accurate spatial relationship. Paintings took on a flat appearance.
At the core of the new perspective thinking was the idea of projecting a three-dimensional subject onto a two-dimensional plane. This was first done optically with perspective devices that used a transparent plane. (fig. B) The artist viewed through a peephole to essentially trace the subject. Other methods used a device known as camera obscura that used a lens to project the image onto a drawing surface allowing the artist to outline the subject. While these techniques were used well into the renaissance, the artist still was limited to render only what was physically available in front of him.
In time, mathematical and geometrical relationships began to be understood that made these physical devices obsolete. Artists could render correct visual representations of perspective without tracing. Foreshortening could be calculated and predicted. It was understood why shadows from the sun were different than shadows from a candle. With only a piece of paper an artist could work out real-world optics to describe size, placement, shape, angles and foreshortening that matched reality. Objective formulas rather than subjective placements gave rise to unparalleled realism in painting. It changed the way people viewed the world and forever improved art.
Visual effects application
As a compositor, I am often asked to shoot elements to match existing footage or stills. Unlike shooting all original elements where camera placement, lens, tilt angle, etc. can be logged and reset in multiple locations. Existing footage and stills most often do not come with camera data. Using a few perspective tools to deconstruct footage can provide a close approximation of camera placement, framing and focal length on set.
A visual effects shot of a car driving into a restaurant at dusk in a desert setting seemed simple enough, and it was. (fig. C) In reality the restaurant did not exist in the desert but was an existing still photo shot in another state. The cars were to be shot over green on location in the desert as was the desert floor and sky at dusk. Creating the desert floor and sky was straightforward and admittedly very forgivable in the perspective realm. However, matching the angle of the driving car and the parked cars was crucial. A camera too high or low, or a lens too wide or narrow would give poor results compromising the believability of the shot.
Before shooting could begin, a single photo of the restaurant was selected. The agency supplied a number of photos, each shot with various lenses from various angles. A commitment was necessary as the whole shot would be built off this single still image. In essence, the perspective within the photo would provide the information needed to expand the scene once we were on location.
A perspective breakdown was done of the selected restaurant. The still photo was a two-point perspective angle. Most instances where a corner is closer to camera than any other flat plain will be two-point perspective. Here a set of converging lines to the left and right help us determine that it is a two point perspective shot.
When working out perspective details many variables build on each other not unlike mathematical calculations. The first variables needed are the two vanishing points (fig. D).
Starting with the right side vanishing point (RVP): In this case the gutters along the right side give an obvious line. The gutter line is continued off the frame pictured here in green.
At least two lines must intersect to find the actual vanishing point. Extending the chimneystack line and the gutter line give a useful point. However, a more accurate point can be determined by locating a line further away. In this case, a line down low such as the sidewalk is used. Still another line is pulled from the dormer roof on the left side of frame. The dormer roofline intersects at the same point as the previous lines. Lines could have also been pulled from the clapboards and windows as well.
In a perfect world where these lines are easily seen and objects are square, two lines only are needed to establish a vanishing point. Having never been to the location presumptions must be made. Therefore, multiple perspective lines showing a pattern of agreement are used. If multiple lines converge on a single point we have to assume the presumptions were accurate.
Establishing a left side vanishing point (LVP) is done in a similar way. Often there are not straight lines that are apparent. However, observations can be made to find lines, especially in man-made structures. In this example, the two dormer peeks are presumed to be in the same height and plane as the side of the building.
A line is pulled from the peek of each connecting the two points extending the line.
Another line is pulled from the cross-brace atop the entrance. A low line is pulled from the base of the stack-stone columns. The sidewalk appears to slope down as it comes toward camera so a line here will not give accurate results.
Once two vanishing points are established a line can be drawn between them. (Pictured in red, fig. E) This is the horizon line and will be the basis for a number of calculations. Another name for the horizon line is the eye-height, or in our world, camera height may be more relevant.
This is the level of the camera in this setting.
Often, I think of the horizon line as a plane instead of a line. The camera is looking down the knife-edge of a plane that cuts the camera lens in half top to bottom. Every object the plane intersects is at the same height including the camera. It makes no difference if the horizon line is high or low in the given frame. The position of the horizon line relative to the frame merely indicates the tilt up or down of the camera. So, the camera is viewing down on everything below the horizon line and conversely, the camera is viewing up at everything above the horizon line.
Determining camera height
Camera height is important because the additional elements need to be within the same dimensional space and viewed from the same angle. A camera too high or low would cause the car to seem off. Calculations to determine camera height can (and should) be done before arriving on set. Too, I find it useful to do before shooting begins, as it is one of the most commonly mistaken variables.
To determine camera height the final composite shot needs to be considered. It was determined the curb from the still photo would not be included in the final composite. The curb effectively raises the restaurant. Working with agency creatives it was decided the desert floor would be level with the bottom of the front door. Knowing this, calculations about camera height would consider the restaurant entrance to be the plane upon which the rest of our world would exist.
With this information an accurate approximation can be made for the camera height by determining the distance from the horizon line (red line) to a horizontal plane that is the floor of the restaurant. In this case the horizon line (red line) to the base of the stacked stoned columns.
Measurements were not supplied with the still photo, but approximations can be made. The bench to the right of the entrance works well for these purposes. (fig. F) The top of the chair reaches the horizon line (red line) and the chair sits on the floor of the entrance to the restaurant. Most furniture complies with standard measurements within a few inches so approximations can be made.
Knowing this, similar chairs were located, measured and found to be 34” (86cm). If more precision was needed, I could have called the restaurant and asked a manager to measure the chair height. If the chair had not been in the photo, a measurement could be taken from the door height. Then calculations could be made to determine the fraction of the door that represents the height to the horizon. Inserting the value of 34” into our bracketed measurements for the door indicated the door to be 93.5” (7’ 9.5”). Most commercial doors in the US are between 7’ and 8’. The cross-brace above the stacked stone pillars was determined to be 136” (11’ 4”).
All of these measurements seem to fit sizes in the real world. None of the measurements calculated seem out of reason. For example, a chair that was 48” high may not seem wrong but a doorway that was 11’ high probably would. Multiple measurements are done to find patterns of agreement. The world makes sense and sizes seem to be based upon reality. This helps to ensure that our assumptions are correct or at least within reasonable tolerances.
Lastly, a graphic of a 6’ man was placed in the scene and sized based on the 34” scale. Again, the size of a human seems to agree with the scale.
A mockup drawing of the restaurant (sized down/ repositioned) with the location of the car was created. (fig. G) This is a layout of the final framing (in blue) of elements and was discussed with creatives for composition and approval. The car was drawn as a block (in red) using the same perspective information determined from the still photo. In this case it was determined to position the car straight in line with the restaurant; in other words the car would be parallel and perpendicular to the restaurant. Other options with the car positioned at angles were also offered and declined.
The elements of the driving car, parked truck, desert floor and sky were shot on location in the desert. The car and truck were shot over green screen as elements to be composited over the desert floor and restaurant. A number of variables would be determined on set; namely pan, tilt, distance from the car and lens. Any one of an assortment of prime lenses was possible.
The day before, late day winds caused problems on set, so shooting the plates would happen mid day instead of dusk as initially planned. Because of the mid day sun, hard shadows became a problem but were minimized by a silk hung overhead. Because the green screen was so large, once set it would not be moved. The camera and car would be positioned to set the shot.
A common mistake is to set a camera at a random height, then tilt up or down to match a horizon line of an existing photo. But of course tilting to match a horizon line is not matching camera height. The height of the camera will have a dramatic effect on perspective.
It is important to observe that objects intersecting the horizon line are all the same height. (fig. G) Given the red horizon line on the mockup, it was determined that the camera (if placed on a level surface with the front door of the restaurant) should be 34” high. So anything that is 34” high will be even with the real horizon. The measurement I had received for the height of the car body was 32”, so if the camera is positioned correctly, the horizon and car body will be close in height.
It is important to note that the objective is to match the lens to the mockup drawing not the original still photo. Once the still photo was sized smaller, the lens had effectively changed to a wider angle. The new framing in the mockup drawing was a wider field of view because it encompassed the original photo.
To match the lens to the mockup drawing a camera with a 24mm lens was placed on set at 32” height. The camera was tilted to match the horizon as indicated by the mockup drawing (approximately 1/3 to the bottom of frame). The camera was positioned front to back on the head so that the nodal point was centered on the pan axis allowing the image plane to rotate on its own axis. The 24mm lens was chosen to begin work as a ‘best guess’. A couple of exercises would indicate if a wider or narrower lens would be required to match the mock up.
In the realm of perspective, the lens can be thought of as angle of view; the amount of the world the viewer (or lens) is able to view limited by the frame. A wide angle of view will push the vanishing points closer together as the viewable area is greater. A narrow angle of view will push the vanishing points further apart. Noting the relatively close vanishing points in the mockup drawing it was clear that the lens would be wide angle. I demonstrated the concept of angle of view (lens) vs. vanishing point distances with an apple box and an SLR. First, the apple box was shot with a wide 20mm lens. The camera was 1’ from the box. (fig. H) Notice the vanishing points are just outside of frame. Stepping back 6’ I shot the same apple box with a 70mm lens. (fig.I) The vanishing points are moved significantly farther outside the frame.
Working in the desert provided several benefits, one being the real horizon line was easily viewed through the camera lens. If the real horizon had not been visible, setting the camera to the 32” height and having someone at a distance in front of the camera with a tape measure marked at 32” would have indicated the horizon to frame. In an environment where level ground is not possible, a laser level would work well.
The car was positioned in front of the camera approximating the size and angle as it appeared in the mockup drawing. Included in the mockup drawing a quarter grid was overlaid to determine size in frame. Considering the mockup drawing, the left vanishing point is positioned less than a quarter frame lengths outside the frame to the left.
By pulling a tape measure from the nearest corner of the car (the bumper) and the furthest corner of the car (the far side of the bumper) the real vanishing point for that lens can be determined by noting the point in which identical numbers match vertically. (The concept above is demonstrated with an apple box set on a similar angle as the car. (fig. J below)
The camera is then panned left. Notice the point at witch the numbers are lined up vertically. In this example it is 10 ½ “. The dark blue line indicates where the real vanishing point is for this object viewed by this lens when intersected by the horizon. In this case, only the left/ right distances are required so horizon line is not included. The pale blue lines are merely to show that identical numbers are not vertical)
The exercise is repeated on the right side of frame. Once the two vanishing points are determined, the distance between them can be compared to the mockup drawing. In this case the 24mm lens was too narrow because the vanishing points were farther apart than the mockup drawing indicated they should be. The mockup drawing indicated the left VP should be less than 1/4 frame lengths outside the left of frame. And the right VP should be a little more than 1/3rd outside of frame. (fig. K below)
The exercise was repeated with an 18mm lens. Camera was moved closer to the car as changing the lens affected the car size in frame. The exercise above was repeated to find left and right vanishing points then compared to the mockup drawing with more favorable results. In this case the 18mm lens was a slightly wider than required.
Tilt / pan
Once the shot is set (fig. L below), tilt and pan can be very forgiving. Since the shot is over green, the top and right sides of frame will be garbage matted out. The shot could have been tilted farther down (raising the horizon line) in order to give more area in post. Also, pan could have been farther left to give more area to work with in post as well. However, if the lens has a great deal of distortion, then it is always best to match the tilt and pan. Using the 18mm lens there was a concern that there would be some distortion so out of caution we framed very close to the mockup drawing.
Because prime lenses were used, it was a matter of getting as close a possible then making small adjustments in post. A slight sizing up the car element and adjusting the final framing a little tighter made both elements fit.
Concern over evening windstorms the day before meant a last minute decision to move the shoot time for these elements from dusk to midday. There was a fear that the large green screens blowing in the desert winds would pose a danger and be impossible to control.
Shooting in the mid-day sun created hard shadow lines of the silk frame across the top of the car as it drove into position. The late decision did not allow enough time to deliver more silks to our remote location. The idea of using no silk was offered but I did not care for the harsh lighting and shadows. Little could be done in post to take away hard lighting. So a single silk with the hard edge problem was shot. This was a calculated risk and a known issue the day of the shoot. The decision created a lot of cleanup on the car and the ground shadows. Given that we had only one silk, I still prefer the look in the final composite to the hard light.
Incidentally, it never got windy.
Epilog: composition and vanishing points
For simplicity, the car was positioned in line with the restaurant sharing the same vanishing points. The client liked the composition and it was a strong image. As an alternate, the car was angled to the restaurant (shown below). In this case the VP is still on the horizon line but not shared with the restaurant as the two share no parallel lines. If it had been decided to bring the car in at an angle to the restaurant, the vanishing points of the car (in green, see right figure M) would have shifted left but would have remained on the horizon line and remained at the same distance from each other. Notice the red vanishing points of the restaurant and the green vanishing points of the angled car are at equal distance. Vanishing points are merely the extrusion of parallel lines of an object that converge at a single point in the distance. Objects that are level to the ground will have VP’s on a horizon line. Objects that rotate and remain level to the ground will shift vanishing points left or right, but the points will remain on the horizon line. Only if an object rotates up or down, as if on an incline would the VP be above or below the horizon.
Keith Adams is owner and visual effects supervisor of SPIN VFX Atlanta.
Hailing from a traditional fine arts background, Adams began his effects career in 1984 as art director for Cinetron Computer Systems, a motion control film studio. He embraced computer graphics as 3D and compositing systems were in their infancy. As an early adopter of AutoDesk’s Flame & Inferno, Adams freelanced throughout the world for several years until 1997, a partnership with SPIN Productions in Toronto was established as he set up shop in Atlanta. Today, Adams works with studios and production companies throughout the country on features, episodic and commercial work.