4.5  Experimental results on rendering quality and evaluation

This section evaluates various combinations of the algorithms presented in this chapter, ranging from rendering to padding and occlusion handling. In the experiments, we have explored the following image rendering algorithms: (1) 3D image warping, (2) mesh-based rendering algorithm, (3) relief texture mapping, (4) rendering using inverse mapping, and (5) rendering using inverse mapping of two source images.

To measure the quality of each rendering technique, a synthetic image is rendered at the same location and orientation of an arbitrarily selected camera (reference view). By comparing the synthetic and captured images, a distortion measure, e.g., PSNRrs can be calculated. As defined in Chapter 3, the PSNRrs distortion metric between a synthetic image Is and a reference image Ir is calculated by

                  (  2552 )
PSNRrs  = 10⋅log10  ------- ,

where the Mean Squared Error (MSErs) is computed by the following equation:

             ∑w  ∑h
MSErs =  -1---      ||Ir(i,j)- Is(i,j)||2,
         w ⋅h i=1 j=1

with w and h corresponding to the width and the height of the image, respectively. For evaluating the performance of the rendering algorithms, experiments were carried out using the “Ballet” and “Breakdancers” texture and depth sequences (see Appendix 9.2). Camera 2 was selected as a reference view while the Cameras 1 and 3 indicate the position and orientation of the rendered virtual views. The measured rendering qualities are summarized in Table 4.1. In the following discussion about the results of Table 4.1, the 3D image warping rendering is used as a reference for comparing results.


(1) 3D image warping

28.1 dB 31.3 dB

warping (S 4.2.1)

padding (S 4.4.2)

(2) mesh-based (S 4.2.2)

28.7 dB 32.3 dB

(3) relief texture

31.9 dB 33.2 dB

warping (S 4.2.3)

image pre-processing (S 4.4.2)

(4) inverse mapping

31.1 dB 32.5 dB

padding (S 4.4.1)

(5) inverse mapping (2 source images)

34.0 dB 35.1 dB

padding (S 4.4.1)

Table 4.1 Rendering quality of various image rendering algorithms. For each complete algorithm, the key steps of the processing are indicated in top-to-bottom order, with references to the section describing the algorithmic steps. The principal step of the algorithm (and thus its name), is printed in bold text.

First, experiments have revealed that a mesh-based rendering method yields a rendering-quality improvement of 0.6 dB to 1.0 dB when compared with 3D image warping. This result demonstrates that a sub-pixel rendering algorithm is necessary to render high-quality images. However, we have found that this approach does not adequately handle occluded regions. Specifically, a mesh-based rendering method interpolates occluded pixels from the foreground and background pixels, so that occluded pixels are padded with a blend of foreground and background colors. Concluding, the gain of the mesh-based algorithm is deteriorated by an inappropriate handling of the occluded regions.

Next, we have compared the rendering quality of the relief texture mapping with the 3D image warping algorithm. Note that in this experiment, the source image is pre-processed using the technique described in Section 4.4.2 and that occluded pixels are padded using the color of the background pixels. Objective rendering-quality measurements show that padding of occluded pixels by the background color produces a noticeable improvement of the rendering quality. Specifically, we have obtained an image-quality improvement of 1.9 dB and 3.8 dB, for the “Breakdancers” and “Ballet” sequences, respectively. Additionally, subjective evaluations demonstrate that the proposed rendering method enables high-quality rendered images. For example, the occluded regions at the right side of foreground characters are correctly extrapolated from the background color and any rendering artifacts are hardly perceived (see Figure 4.12(b) and Figure 4.14(b)). This confirms that deriving the color of occluded regions from neighboring background pixels is a simple heuristic algorithm but an efficient approach.

Finally, the performances of two inverse mapping rendering techniques are evaluated. First, experimental results show that the inverse mapping rendering technique improves the rendering quality up to 3 dB, when compared to 3D image warping. In addition, it can be noted that the relief texture rendering method slightly outperforms the inverse mapping rendering technique. Such a result simply emphasizes that the occlusion-handling technique has a significant impact on the final rendering quality. In this specific case, the occlusion-handling technique, which includes the image pre-processing step, produces significant rendering improvements (see Figure 4.13(a) and Figure 4.15(a)).

Second, the two-image inverse mapping technique is compared with the 3D image warping algorithm. Objective rendering-quality measurements show that a significant rendering-quality improvement can be obtained by combining two source images for synthesizing a single image. For example, an improvement of 3.8 dB and 5.9 dB is obtained for the “Breakdancers” and “Ballet” sequences, respectively. Additionally, subjective evaluations show that occluded regions are adequately handled. For example, the occluded regions at the right of the foreground persons in Figure 4.13(b) and Figure 4.15(b) are correctly defined and rendered accurately. Note that, as opposed to the heuristic padding technique, the occluded region is now correctly defined.

For comparison purposes, we have developed a C++ software implementation of the rendering algorithms. Since the presented work focuses on the quality of the rendered images, the software implementation was not optimized for fast rendering. Practically, our software implementation yields a rendering time that ranges between 3.5 and 7 seconds for each frame (of size 1024 × 768 pixels). Again, note that this rendering time is not obtained from an optimized software implementation. It is thus possible to significantly reduce the rendering time by using a Single Instructions Multiple Data (SIMD) processor architecture, which is available in modern CPU’s. Considering a possible GPU-based optimization of the relief texture mapping algorithm, we have found that using a GPU-based acceleration does not bring execution performance improvements. Specifically, although the execution of the homography transform is performed efficiently by the GPU, the speed for transferring the resulting synthetic image from the graphic GPU memory to the main CPU memory is low, so that not computation but bandwidth is the bottleneck. Currently, this communication bottleneck has been addressed by the latest generation of modern GPU’s.