35.6 Automatic aesthetic lighting⧉
Good lighting is the difference between a snapshot and a portrait, and it is the hardest thing for an amateur to get right. The reason an on-camera flash looks bad is geometric: it is small (so its light is hard, casting sharp shadows) and it sits on the lens axis (so its light is flat, killing the shape-revealing gradient across a face). Professionals fix this physically: they make the source large and put it off to the side, bouncing the flash off a wall or ceiling, floating a softbox on a stand, or stationing an assistant to hold a light exactly where it flatters. That expertise, where to put the light, is what this chapter automates. The move is to treat the light as an actuator under a feedback loop: define an aesthetic objective (soft, directional fill; a clean rim), measure how close the current lighting comes to it, and command the light to move until it gets there. The optimized variable is now the light's aim or position in space, not its pattern or timing, and the loop closes around a goal that is about how the picture looks, not about reconstructing the scene.
35.6.1 Computational bounce flash⧉
Photographers have long known the trick of bounce flash: instead of firing the flash straight at the subject, aim it at a wall or ceiling, so the light arrives at the subject from a large, diffuse, off-axis patch. It is soft and directional, the opposite of harsh frontal flash. The catch is that which surface to bounce off is a real skill: bounce off the wrong patch and the light comes from a bad angle, or there is no suitable surface and the bounce fails. Murmann, Davis, Kautz and Durand (2016) automated the choice. Their computational bounce flash mounts the flash on a motorized head that can rotate to aim anywhere in the room, and the system computes the best bounce patch before firing (Figure 35.6.1).
The procedure: probe the room (the device can fire quick test flashes and observe how light returns from candidate directions, and reason about the room geometry), then evaluate each candidate wall or ceiling patch by the quality of light it would deliver to the subject. A large, off-axis patch that produces soft, directional illumination scores well; a patch that would yield flat frontal light or harsh shadows scores poorly. The motor then aims the flash at the winning patch and fires. The system can choose softer, directional bounce lighting automatically in suitable rooms. Wall color, available surfaces, flash power, subject motion, and the chosen aesthetic objective constrain the result. For this part: the bounce direction is the optimized degree of freedom, and the objective is explicitly aesthetic (the flattering quality of light on a face), not a physical reconstruction.
35.6.2 Drone lighting: flying the light into place⧉
The bounce flash steers an aim; the next system steers a position in three-dimensional space. Srikanth, Bala and Durand (2014) mounted a light on a quadrotor and let it fly itself into the right place to produce a specific lighting effect, rim light, the bright contour that separates a subject from the background and gives a portrait depth and polish. Rim light is notoriously fiddly to set up by hand because it requires a light placed behind and to the side of the subject, aimed back toward the camera, and it has to be re-placed every time the subject or photographer moves. So make the light an autonomous actuator: a feedback loop measures the rim in the camera's live image, specifically the width and placement of the bright edge along the subject's silhouette, compares it to the photographer's target rim, and commands the drone to reposition so as to drive the measured rim toward the target (Figure 35.6.2). As the subject turns or the photographer walks around, the loop continuously re-flies the light to hold the desired rim. The light source has become a flying robot servoed to an image-based aesthetic objective.
Here the optimized degree of freedom is the light's position in space, and the controller is a closed loop around a measured look. Bounce flash and drone lighting differ in what they actuate, a motorized aim versus a flying position, but they are the same idea: hand the placement of the light to a feedback loop that measures the desired aesthetic and drives a physical light source to achieve it. They also gesture at the broader active-capture / robotic-photography theme (a camera or its lights as actuated, goal-seeking instruments) that recurs in the FUNDAMENTALS treatment of cameras beyond photography.
