Digital photo stitching application is the workhorse in the panorama-making process, and may range from providing a completely automatic one-click stitching, to some additional time-consuming manual process. This can be part 2 of your tutorial, which assumes all individual photos have been properly captured (stage 1 below is finished); for stage 1 and a summary of the complete stitching process kindly visit part 1 with this tutorial on digital panoramas.
As a way to begin processing our combination of photos, we need to select an appropriate computer software. The most significant distinction between options is within the way that they opt to address the tradeoff between automation and adaptability. Generally, fully customized stitching software will always achieve better quality than automated packages, but it can possibly result in being overly technical or time consuming.
This tutorial aims to improve knowledge of most software stitching concepts by keeping the discussion as generic as you can, however actual software features may reference a treatment program called PTAssembler or PTGui (front-end for PanoTools or PTMender). PTAssembler incorporates a fully-automated one-click stitching option, as well as providing for nearly all possible custom stitching possibilities in other programs.
During the time of this post, other notable programs include the ones that come packaged with the camera, like Canon PhotoStitch, or popular commercial packages for example Autostitch, photoshop software, Arc Soft Panorama Maker, Panorama Factory and PanaVue, among others.
STAGE 2: CONTROL POINTS & PHOTO ALIGNMENT
Panorama stitching software uses pairs of control points to specify areas of two camera photos that make reference to a similar reason for space. Pairs of control points can be manually selected by visual inspection, or these could be generated automatically using sophisticated matching algorithms (like Autopano for PTAssembler). With many photographs, greatest results could only be accomplished with manual control point selection (which is often the most time-consuming stage in the software stitching process).
The example above shows a selection of four pairs of control points, for 2 photos in just a panorama. The very best control points are the type which can be in relation to highly rigid objects with sharp edges or fine detail, and are spaced evenly and broadly across each overlap region (with 3-5 points for each overlap). This means that basing control points on tree limbs, clouds or water is ill-advised except when absolutely necessary. It is actually because of this recommended to continually capture some land (or other rigid objects) inside the overlap region between all pairs of photographs, otherwise control point selection may prove difficult and inaccurate (like for panoramas containing all sky or water).
The example below demonstrates a situation where only detailed, rigid section of each image is incorporated in the silhouette of land on the very bottom-thereby making it hard to space the control points evenly across each photo’s overlap region. During these situations automated control point selection may prove better.
PTAssembler carries a feature called “automatically micro-position control points,” which operates by with your selection as being an initial guess, then planning to all adjacent pixels in a specified distance (including 5 pixels) to find out if they are a greater match. When stitching difficult cloud scenes like that shown above, this effectively combines the advantages of manual control point selection with the ones from automated algorithms.
Another consideration is just how far awayfrom the camera each control point is physically located. For panoramas taken without a panoramic head, parallax error may become large in foreground objects, therefore better results is possible by only basing these on distant objects. Any parallax error within the near foreground may not be visible if each of these foreground elements usually are not contained throughout the overlap between photos.
STAGE 3: VANISHING POINT PERSPECTIVE
Most photo stitching software gives the capability to specify the location where the reference or vanishing point of perspective is located, along with the sort of image projection.
Careful collection of this vanishing point might help avoid converging vertical lines (which will otherwise run parallel), or even a curved horizon. The vanishing point is normally where one could be directly facing if they were standing inside the panoramic scene. For architectural stitches, such as the example below (120° crop through the rectilinear projection), this point can also be clearly apparent by simply following lines into the distance which can be parallel to one’s brand of site.
Incorrect placement of the vanishing point causes lines laying from the planes perpendicular towards the viewer’s line of site to converge (even though these would otherwise appear as being parallel). This effect may also be observed simply by using a wide angle lens within an architectural photo and pointing the camera significantly below or above the horizon- thereby giving the impression of buildings that are leaning.
The vanishing point is additionally critical in very wide angle, cylindrical projection panoramas (such as the 360 degree image shown below). It might exhibit different looking distortion if misplaced, producing a curved horizon.
In the event the vanishing point were placed too high, the horizon curvature would be from the opposite direction. Sometimes it might be difficult to locate the particular horizon, as a result of presence of hills, mountains, trees or any other obstructions. For such difficult scenarios the location of the horizon could then be inferred by placing it with a height which minimizes any curvature.
Panorama stitching software also often gives the option to tilt the imaginary horizon. This is very useful as soon as the photo containing the vanishing point was not taken perfectly level. For this scenario, even if the vanishing point is put in the correct height, the horizon might be rendered as having an S-curve in the event the imaginary horizon fails to align with the actual horizon (in the individual photo).
If the panorama itself were taken level, then this straightest horizon would be the one which yields a stitched image whose vertical dimension is the shortest (which is a method sometimes employed by stitching software).
STAGE 4: OPTIMIZING PHOTO POSITIONS
After the control points, vanishing point perspective and image projection supply been chosen, the photo stitching software may then start to distort and align each image to produce the ultimate stitched photograph. This can be the most computationally intensive step at the same time. It functions by systematically searching through mixtures of yaw, pitch and roll as a way to minimize the aggregate error between all pairs of control points. This process can also adjust lens distortion parameters, if unknown.
Keep in mind that these photos are slightly distorted; this really is to emphasize that when the stitching software positions each image it adjusts for perspective, and that the level of perspective distortion is determined by that image’s location relative to the vanishing point.
The important thing quality metric to pay attention to may be the average distance between control points. If this distance is large in accordance with paper size, then seams can be visible regardless of how well these are generally blended. One thing to check is whether or not any control points were mistakenly placed, and they follow the other guidelines listed in stage 2. When the average distance is still too large then this may be a result of improperly captured images, including parallax error from camera movement or perhaps not using a panoramic head.
STAGE 5: MANUALLY REDIRECTING & BLENDING SEAMS
Ideally one may want to position the photo seams along unimportant or natural break points within the scene. When the stitching software supports layered output one could perform this manually utilizing a mask in photoshop:
Without Blend Manual Blend Mask from Manual Blend
Note how the above manual blend evens the skies and avoids visible jumps along geometrically prominent architectural lines, like the crescent of pillars, foreground row of statues and distant white building.
Ensure that you blend the mask over large distances for smooth textures, such as the sky region above. For fine detail, blending over large distances can blur the photo when there is any misalignment between photos. It is therefore wise to blend fine details over short distances using seams which avoid any easily noticeable discontinuities (look at the “mask from manual blend” above to discover exactly how the sky and buildings were blended).
Alternatively, manually blending seams could become extremely time intensive. Fortunately photo making software has an automated feature which could perform this simultaneously, as described over the following section.
STAGE 5: AUTOMATICALLY REDIRECTING & BLENDING SEAMS
Among the best strategies to blend seams in the stitched photograph is to use an approach called “multi-resolution splines”, that may often rectify even poorly captured panoramas or mosaics. It works by breaking each image up into several components, just like how an RGB photo may be separated into individual red, green and blue channels, although in this case each component represents a different scale of image texture. Small-scale features (including foliage or fine grass) have a superior spatial resolution, whereas larger scale features (such as a clear sky gradient) are said to obtain low spatial resolutions.