3D image synthesis from depth encoded source view

[0001] This application is based upon and claims the benefit of priority under 35 U.S.C. § 119 from the prior Australian Patent Application No. 2002952874, filed Nov. 25, 2002, the entire contents of which are incorporated herein by reference.

[0002] The present invention is directed towards an improved technique for the generation of images for use with autostereoscopic displays. In particular, the present invention relates to an improved method of producing images derived from a 2D image and associated depth map.

[0003] A number of autostereoscopic displays are starting to appear on the market, some of which require multiple images in order to provide an autostereoscopic image, and enable the viewer to retain a stereoscopic effect despite movement of their head.

[0004] Such displays generally require images comprising a number of views created from a number of laterally displaced cameras. Such views can be originated from real cameras, generated using computer graphic image (CGI) techniques, or synthesized from a 2D image and an associated depth map.

[0005] The synthesis of multiple images from a 2D image and an associated depth map has previously been disclosed by the present Applicants in Australian patents AUS 714759 (10884/97) and AUS 738692 (16472/99) included here in full by reference.

[0006] In U.S. Pat. Nos. 6,118,584 and 6,064,424, included here in full by reference, van Berkel describes an autostereoscopic display that requires seven views. In German patent PCT WO 01/56302 A1, included here in full by reference, Grasnik describes an autostereoscopic display that requires eight views.

[0007] To those skilled in the art such displays are known to require multi-views or integer multi-views in order to display autostereoscopic images.

[0008] Given the commercial availability of such multi-view displays there is a corresponding requirement for suitable images or content.

[0009] There is thus a need for a more efficient technique for developing high quality images for multi-view autostereoscopic displays. Ideally the technique will be generic and can be applied to any display requiring multiple views (e.g. lenticular, parallax barrier, wavelength selective filter).

[0010] It is therefore an object of the present invention to provide an improved method for the real time generation of images, suitable for use with multi-view autostereoscopic displays.

[0011] With the above object in mind the present invention provides a method of creating images suitable for use with a multi-view autostereoscopic display including the steps of:

• [0012] 1. A view calculation means for determining a fractional view for a given position of the synthesized 3D image.
• [0013] 2. A source pixel calculation means for relating the calculated view number to a position in the source image.
• [0014] 3. A weighted average means for calculating a source pixel value given a fractional source position.

[0015] In another aspect the present invention provides a method for generating a synthesized 3D image from a source image including the steps of:

[0016] determining a fractional view number for each pixel of a pixel array, and

[0017] referring to a depth map of said source image to determine a location offset of each pixel.

[0025] Whilst content can be obtained from a number of laterally displaced cameras this invention relates to the generation of images synthesized from a 2D image and an associated depth map.

[0026] In the preferred embodiment of the present invention the process of generating a synthesized 3D image includes two steps: namely the mapping of a position on the 3D display device to a fractional view number and the subsequent determination of a suitable source pixel to represent the said fractional view number.

[0027] A “view” may be defined as the projection of a 3D object or scene on to a 2D image. For the purposes of multiview displays views are generally converged on a point of focus and numbered from left to right as shown in FIG. 1. The optimal configuration of views depends on the characteristics of the screen and the contents of the scene. Each view is captured by some form of camera or capture device, whether physical or virtual. FIG. 2 illustrates two views of the object in FIG. 1—the shift in perspective effectively encodes the 3D structure of the scene.

[0028] An alternative definition of a view is to consider the angle between two cameras and a fixed point in the scene. For example, in FIG. 1 view 3 may be considered to be at an angle of 2 degrees from the front-on view 4. The present invention is not restricted to a fixed configuration of views but is able to generate arbitrary fractional views from a source view and an associated depth map (as described in the Applicants previous disclosure PCT/AU96/00820).

View Determination

[0029] The first step of generating a synthesized 3D image involves determining a fractional view number for each pixel of the pixel array. FIG. 3 illustrates a small section 5 of a typical pixel array. The array 5 has a number of cells or pixels each containing a color and intensity value. Overlying the pixel array is an optical component 6 which may be arranged at an angle 7 α to the pixel array.

[0030] A generic formula for determining the source view for any given pixel of the array was previously described by Van Berkel as

[0031] $N=\frac{\left(k+k_{\mathrm{offset}}-3l\tan \alpha \right)\mathrm{mod}X}{X}{N}_{\mathrm{tot}}$
where k is horizontal pixel index

[0032] k_offset is horizontal shift of the lenticular lens array relative to the pixel array

[0033] α is angle of lenticular lens array relative to the pixel array

[0034] X is the number of views per lenticule

[0035] N_tot is total number of views

[0036] N is the view number of each sub pixel k,l

[0037] Given characteristics of the screen such as the angle 7 between the lenticular and the pixel array α as well as the pitch of the lenticules a source view (N) can be calculated for any position (k,l) in the synthesized 3D image.

[0038] Although Van Berkel's formula describes view determination for lenticular displays it may also be applied to parallax barrier systems. The barrier will simply obscure some views. Similarly, the formula may also be applied to wavelength selective filter arrays in which case some views will be filtered. Further although van Berkel teaches taking the nearest integer view (N) for the actual mapping the fractional result from the formula may be used directly.

[0039] In practice, the optical component overlying the pixel array gives rise to a repeating pattern which maps view numbers to the pixel array. This pattern can be represented as a mask, which is repeatedly applied. This leads to a gain in efficiency as van Berkel's formula only needs to be calculated once for each mask element.

View Generation

[0040] Once a fractional view, N, has been calculated for a given k,l position in the pixel array of the synthesized 3D image the present invention provides an inventive means for generating a pixel for the required view. This approach uses a depth map to determine a location offset of the pixel in a single source view. This offset effectively simulates a perspective transformation by introducing parallax.

[0041] FIG. 4 illustrates the pixels 8 of the source view corresponding to the synthesized pixel array in FIG. 3. Similarly, FIG. 5 illustrates an array of depth values 9 corresponding to the pixels of the source view. To calculate the position of the pixel in the source view corresponding to the required view, N the following formula is used:
k′=k+3(Z_k,lP_shift(N)P_strength−P_offset)
where:

• • [0042] Z_k,l is the depth at position k,l (which varies between 0 and 1)
  • [0043] P_shift(N) is a function relating the view N to a parallax shift (−1 . . . 1)
  • [0044] P_strength controls the total parallax strength
  • [0045] P_offset is used to control the parallax offset (or false parallax)

[0046] The function describing the relationship between a view N and the associated parallax shift P_shift(N) depends on the optical characteristics of the 3D screen in question.

[0047] FIG. 6 shows a realistic example of the P_shift(N) function for a seven view display. With reference to FIG. 1 it is apparent that the shift is greater at the extreme views (1 and 7) than in the central view (4). It should be noted that this formulation can be used to relate shift to fractional views. For example, in FIG. 3 view 3.127 can be related to a shift of approximately 0.25. In practice the P_shift(N) function may not be linear—for example it may be desired to reduce the amount of shift towards the extreme views.

[0048] The calculated position k′, l′ may not be an integer, in which case the nearest position may be used. Alternatively, the nearest integer source pixels may be averaged using a weighted average calculation. As an alternative to weighted average calculation, any statistic of the closest pixels to k′ may be calculated.

[0049] The current invention improves on the prior art by removing the need to completely generate multiple views and is therefore significantly more efficient.

[0050] For example, when synthesizing a 3D image using nine views the 3D image only contains 1/9 of each view—the remaining 8/9 are discarded. With the direct source and depth synthesis technique we only generate the 1/9 of each view we require, effectively reducing the amount of calculations required by nine. In general, for display with N views the current invention requires N times less calculations than the prior art.

[0051] A further aspect of the current invention improves on the prior art by providing a mechanism for generating arbitrary fractional views. The prior art teaches the generation of a synthesized 3D image from a number of fixed views. The said views, are fixed at the point they are captured by a physical or virtual camera. The present invention includes a view generation means, which enables arbitrary views to be generated during image synthesis. This provides improved flexibility as the same data can be used to configure the views for any multiview 3D display.

[0052] The current invention improves on the prior art by providing a more efficient means for synthesizing a 3D image. Traditionally, a 3D image is synthesized by first generating the required number of complete views and then subsampling these views to form a composite 3D image. For example, in an autostereoscopic display requiring 24 views, 24 individual images would have to be generated requiring significant computation and memory resources. The current invention avoids the need for generating these views by synthesizing the image directly using a single view and a depth map.

[0053] An illustrative system for creating images suitable for display on a multi-view autostereoscopic display device is shown in FIG. 7. A view calculation module 72 receives position parameters and determines a fractional view for a given position on the display device. Then a source pixel calculation module 74 maps a source pixel to the position as a function of a depth of the pixel at the position and the fractional view at the position. The fractional view number can be determined by the above described formula for the view number N and the source pixel can be determined by the above formula for k′.

[0054] Whilst the method and apparatus of the present invention has been summarized and explained by illustrative application it will be appreciated by those skilled in the art that many widely varying embodiments and applications arc within the teaching and scope of the present invention, and that the examples presented herein are by way of illustration only and should not be construed as limiting the scope of this invention.