Computer Graphics Laboratory ETH Zurich


Vision-based calibration of parallax barrier displays

N. Ranieri, M. Gross

Proceedings of SPIE 9011 (San Francisco, USA, February 3-6, 2014), pp. 90111D


Static and dynamic parallax barrier displays became very popular over the past years. Especially for single viewer applications like tablets, phones and other hand-held devices, parallax barriers provide a convenient solution to render stereoscopic content. In our work we present a computer vision based calibration approach to relate image layer and barrier layer of parallax barrier displays with unknown display geometry for static or dynamic viewer positions using homographies. We provide the math and methods to compose the required homographies on the fly and present a way to compute the barrier without the need of any iteration. Our GPU implementation is stable and general and can be used to reduce latency and increase refresh rate of existing and upcoming barrier methods.

Figure 1: Illustration of two optimal barrier patterns. On the left side, barrier positions are used to define slits acting as pinhole, exposing left view's pixels (dashed lines) to the left eye Eleft and the right view's pixels (solid lines) to the right eye Eright only. On the right side, the barrier positions define the border of alternating blocking and translucent barrier patches, providing much more brightness. Latter approach can further be improved by swapping the translucent and opaque patches each alternating frame, providing full spatial resolution.


Stereoscopic parallax barrier displays deploy an image layer with two scrambled views and a barrier layer to multiplex the two views to distinct eye positions. A special translucent/opaque pattern is shown on the barrier layer, which exposes pixels of one view to one eye but blocks their sight to the other eye and vice versa. Careful alignment of barrier pattern and scrambled image pattern is required and crucial to avoid crosstalk and aliasing. Furthermore, computation of the barrier pattern in existing methods often relies on known display geometry. Imperfections and misalignment caused by the manufacturing process can thus lead to Moiré patterns, crosstalk and other artefacts. Also, the iterative nature common to these methods induces a growing latency with increasing display resolution.
Therefore, we propose a computer vision based calibration method for static and dynamic parallax barriers. We relate the barrier layer to the image layer using homographies for arbitrary viewing positions. Furthermore, we demonstrate how algorithms for camera calibration can be used to derive the required homography for any given eye position. We provide a method how to use the homographies to compute all barrier positions independently of all others, and thus without the need of any iteration. This method is then used to compute hardware accelerated the whole barrier pattern on GPU, increasing refresh rate and decreasing latency. Proof of concept of our method is given in a physical prototype and we conclude our work with a discussion of our results.

Figure 2: The top row shows our transparent stereoscopic display prototype once without displaying anything to illustrate transparency (top left) and once showing the calibration pattern that was used in our calibration algorithms (top right). A RGB liquid crystal display is used as barrier layer loosing much transparency in the embedded color filters. Using a gray-scale liquid crystal display as deployed in e.g. medical screens would improve transparency significantly. The bottom row shows results captured on our prototype, calibrated with the proposed method. Left eye's view (bottom left) and right eye's view (bottom right) show the image separation. Time-multiplexing for alternating transparent/opaque patches was applied to regain full spatial resolution. Crosstalk at the right border of the display is due to the slightly non-planar projection surface. The strong radial fall-off comes from the projector and is not subject of our work.


To assess quality of our calibration we use the pixel distance between calibrated homography and homography computed by our algorithms as described in Section 6. Sub-pixel accuracy in display pixel space has been achieved for completely planar barrier and image layer. We have observed an increasing error for slightly bent or non-planar layers as they do not fit the assumption of planar surfaces in our algorithm.
Our implementation is able to run at 120Hz for FullHD on a state of the art desktop graphics card, tested with an opaque parallax barrier display. The 60Hz refresh rate (limited by the projector) in our presented prototype could be achieved without any problems. The crucial eigendecomposition can be computed in less than a millisecond due to the small and fixed size of the homography. Barrier computation is as fast as rendering at most 1920 quads on GPU for full HD.
Results captured on our prototype are shown in Figure 2. The two bottom images show left and right eye's view with clear color separation. Slight crosstalk is visible at the right border of the screen, coming from the slightly non-planar back-projection screen.


Download Paper
Download Paper