TW463135B - Method for efficiently encoding three-dimensional graphics models - Google Patents

Abstract

A method of efficiently encoding the geometry of three-dimensional graphics model objects is disclosed. The disclosed method first compares a geometry value from the three-dimensional graphics model with a predicted geometry value that is created from previous geometry predictions. The method then quantizes an error difference between the actual geometry value and the predicted geometry value to produce a quantized error difference. The quantized error difference is then entropy encoded.

Description

46313 5 V. Description of the Invention (l) Field of the Invention The present invention relates to the field of computer graphics. In particular, the present invention discloses a method for efficiently encoding three-dimensional graphics patterns. BACKGROUND OF THE INVENTION 'The level of detail and complexity of 3D graphics has increased dramatically in recent years. Providing complex 3D graphic images requires detailed 3D object patterns. Detailed 3D object patterns include thousands or millions of vertices and orthogonals. The energy to process such a large amount of 3D object model data becomes increasingly difficult. Although the computing power of personal computers (PCs) is almost doubled every two years, the memory bandwidth between the processor and the main memory becomes 7ιαι π π ~, the channel that carries three-dimensional object data from the main memory to the processor Often it is not possible to maintain the data rate required to provide images in real time. Furthermore, switching between main memory and specific graphics controllers is a bottleneck that limits performance. … In the three-dimensional object model with! Angular mesh foundation, each triangular surface is orthogonal to two planes by two apexes and intersects with each other like a month). In many three-dimensional object models (^ 'maternal apex — a positive coordinate stored as four bytes, the apex of a trigonometric triangle and each triangle of the orthogonal network have 72 points. Therefore, the triangle

6 coordinates / top χ3 top / triangle = 72-bit definition (4 bytes / coordinate X action, where each box contains one million _ triangles). To generate real-time seconds, the data rate of the memory bus is angular and the frame rate is 60 frames / second. Such a model ignores jealous surnames with at least 4.3 billion bytes / more memory bus bandwidth and two characteristic coordinates, which may consume the usual 3 memory buses of current personal computer systems. Bandwidth is 500 thousand

Page 5 46313 5 V. Description of the invention (2) Range of ten thousand bytes / second. Obviously, the main bottleneck of the memory bus bandwidth in the three-dimensional graphics perspective ^ In a research report titled "Geometric compression through topological surgery" in 196, 'Taub in and Rossi gnac proposed a Compression and decompression techniques that significantly reduce the size of 3D object models. By compressing the 3D object model, less 3D object model data needs to pass through the limited bandwidth bus between the main memory and the processor. therefore. More complex 3D object models can be routed through the same memory bus. The method introduced by Taub and Rossignac is part of the fourth compression standard proposed by the Moving Picture Experts Group later known as MPEG4. For details, please refer to the document "Description of Core Experiments on 3D Model Coding", which was proposed at the Rome MPEG4 conference in December 1998. It is the title of IS0 / IEC JTC1 / SC29 / WG11 WPEG98 / M4312. SUMMARY OF THE INVENTION A method for encoding a three-dimensional graphics model is disclosed. This method first compares a geometric value with a three-dimensional graphics model with a predicted geometric value. This method then quantifies the error difference between the geometric value and the predicted geometric value to produce a quantized error difference.13 Other objects, features, and advantages of the present invention will become apparent from the accompanying drawings and from the detailed description below. The purpose, features, and advantages of the present invention will become apparent to those skilled in the art from the following detailed description, where: Figure 1 A illustrates the block of a geometric encoder used in the encoding part of the MPEG4 standard.

Page 6 46313 5 V. Description of the invention (3) Figure. Fig. 1 shows a block diagram of a geometric decoder for the coding portion of the 3D model of the MPEG 4 standard. Figure 2A illustrates a block diagram of a geometric encoder built by a professor of the present invention. Figure 2B illustrates a block diagram of a geometric decoder built by a professor of the present invention. FIG. 3A illustrates a wire-frame perspective view of a triangular mesh model of an electric drill. Figure 3 B s Youming is a bit / triangular chart corresponding to the maximum squared quantization error corresponding to the two different geometric codes of the electric drilling machine of Figure 3 A. Fig. 3C illustrates a bit / triangular graph with respect to the mean squared quantization error of the two different geometric codes corresponding to the electric drill of Fig. 3A. Figure 4 A illustrates a wireframe perspective view of a triangular mesh model of a horse. Fig. 4B illustrates a bit / triangular graph with respect to the maximum squared quantization error corresponding to the two different geometric codes of Fig. 4A. Fig. 4C illustrates a bit / triangular graph with respect to the mean squared quantization error of two different geometric codes corresponding to the horse of Fig. 4A. A detailed description of the system example reveals an effective one " 一 娜 .factory ....... The taste of the dust and the device. In the description of the following 3, for the purpose of explanation, specific terms are described to thoroughly understand the present invention. However, these specific details of ^ familiar with the art are not necessary to implement the present invention. The present invention has been referred to the technology proposed as the MPEG4 standard. Take for example '. However, the same technique can be easily applied to other shaped bears, 隹 隹 46313 5 V. Description of the invention (4) Shaped system. Three-object model compression Complicated three-dimensional graphics need to be reported. 2D object models have extremely detailed 2D object models for computer resources. Large 3D object models require a large number of everlasting requirements. In particular, the good main memory comes from fast storage space for storage in perspective, a large amount of streaming bandwidth to move the 3D object model for access, and a large memory pool for complex 3D object models. Deal with long-term processors. To reduce these requirements for bus bandwidth, there are many ways to store this memory and shrink this three-dimensional object model. The three = two-dimensional graphics system has adopted an ideal compression system that can provide a very good compression ratio and the most j-object model.物件 Small object distortion and simple real MPEG4 3D model encoding are known to the animation picture expert group — _ An image compression system known as " MPEG4, 'is a definition of the objects in the image to define the material / The image ^ annoyance is known by the three-dimensional object model. The compressed image can be reconstructed by describing the scene with a three-dimensional object model. In order to efficiently store MPEG4 3D object models, the MpEG4 standard must adopt standard 3D model coding technology, which is generally referred to as 3DMC. One of the 3D model coding (3DMC) technologies proposed to the MPEG4 committee is the model coding system defined by Taubin and Rossignac in their 1996 study entitled 1 'Geometric compression through topological surgery ".

Taubin and Rossignac's 3D model coding technology (see later)

Page 8 4 6313 5 V. Description of the invention (5) '~ ---- 3DMC / TR) contains two main components: the connector coding element and the geometry coding file. Attachment coding elements attempt to reduce the redundant nature of many presentations of polyhedrons or triangles. For example, a triangular mesh—a straight but inefficient technique—represents each triangle in the mesh by directly referring to the three vertices of the triangle. This 3DMC / TR system first decomposes a network into triangles and vertices across trees. These trees are individually encoded in a lossless manner. 3DMC / TR's geometric coding element performs lossy compression to reduce the size of graphics data. Specific% ’This geometric coding element compresses vertex coordinates, vertex parent, color and feature coordinate data. Ben: Ming's focus is on improving the technology of geometric coding of 3D object model data. Therefore, the remainder of this document will focus on how to encode a flat stone horse-° ^ DC / TRt shown) 'to generate a car Λ / Beid point coordinate geometry as a unit cube (Is not Q 1; Hua Yiyang The vertex position of the point in. Consistent scalar quantizer Qt truncates each positive integer n ^ fixed-length integer number i, where L is the quantization number U of each coordinate multiplied by the integer quantization that can be defined by the formula. The number, specifically, the next, 3DMC / TR geometric coding method to steal comparison between other adjacent quantized coordinates. 仃 I vertex coordinates and the geometric distance that usually implies the corresponding vertex ' Coordinate comparison. In particular, in the figure u = near quantized coordinates and the line just across the tree-like adjacent vertex an enter $ tester P to generate from vertex village 'and combine to predict the current vertex position

4 6 3 13 5 V. Description of the invention (6). This method then encodes the error difference en between the predicted value ⑻ and the actual quantized vertex position in. When the vertex coordinates are quantized as a fixed point, the correction vector of these error differences en average has a smaller magnitude than the absolute position. Therefore, these error difference correction vectors can be encoded in a lossless manner with fewer bits. Therefore, this error difference en correction term is further compressed by a consistent encoder 1 4 5 and the standard serial lossless compression technique is used to generate the final series of compression error differences. The 3DMC / TR geometric decoder corresponds to the 3DMC / TR geometric decoder. Turn to the process of the method described in Figure 1A. Figure 1B illustrates a block diagram of a 3DMC / TR geometry decoder. Referring to FIG. 1B, the coherent decoder 155 decompresses the compressed error difference & to generate an uncompressed error signal en. The predictor P then feeds the error signal A into the predicted value record to generate a quantized value. (This quantized value 1n is used by the predictor p to generate further predicted values gj. Finally, ia is dequantized to generate reconstructed geometric values. One of the main disadvantages of the 3DMC / TR geometric coding technique is the direct quantization of the input vertex coordinates. Quite high level of distortion. The distortion of the encoded 3D model introduced by the 3DMC / TR geometric encoder may produce a less acceptable final image. The visual quality of compressed 3D objects is estimated as the maximum squared distortion of two general mining cycles. The quantization error E and the mean squared quantization error '. Assume the original vertex position and W' i is the reconstructed vertex position, where ι = 1, ..., η and η is the total number of vertices. And mean squared quantization error £ _ can be defined as: na

Page 10 46313 5 V. Description of the invention (7)

E

max = SS

Wi-Wi

Wt

Emean = ~ Σ and the dynamic range of the input vertex coordinates is fixed and known), the statistics of the input data are generally unknown. Because designing the best quantizer in the normalized input data is not an option. In the case of morphology, the optimization of the quantizer refers to the smallest flat distortion. If the high-resolution quantizer of the If shape can achieve low distortion and Production; phase: 2. rate. The second disadvantage is the conversion of floating point values to integers; large quantization errors. Xia Huachan-Improved Geometric Encoder / Decoder The present invention introduces an improved encoding technique that provides better results than a 3DMC / TR geometry encoder. The geometric coding method of the present invention quantifies the error difference between the input vertex coordinates and the related prediction values, instead of directly quantifying the input vertex coordinates. The quantified difference is then added back to the same predicted value to form the reconstructed value. This technique is commonly known as differential pulse modulation (DPCM) and is widely used in speech and audio coding. Improved geometric adaptation Figure 2A illustrates a block diagram of a geometric encoder built according to the teaching of the present invention. "Referring to Figure 2A 'a predicted value ϋη minus the input value un generated by the predictor p from a previously reproduced value to form an error signal ~ . The quantizer Qr then converts the error signal

Page 11 46313 5 V. Description of the invention White quantization (8) produces a quantized error signal en. The uniform encoder compresses this quantized error signal to generate a series of compression error signals. An adder generates a reconstructed value ΰη by adding the quantization error signal ′ to the predicted value ϋη. This predictor ρ uses one or more reconstructed values ΰη to form the next predicted value 1. The predictor ρ may have the same form as the predictor P used in the 3DMC / TR geometric coding system. By comparing Fig. 1A with Fig. 2A, it can be found that the original 3DMC / TR geometric encoder can be converted into a DPCM geometric encoder with only a few modifications. The DP CM coding of this form is based on an assumption. If the predicted value (jn is reasonably correct and fairly close to the actual coordinate value Un, the magnitude of the error signal ^ will be small and in most cases is smaller than the actual coordinate value un. In this case, the error signal en changes Will be less than the change in the coordinate value Un. Therefore * quantizing the dynamic range error signal en with the same number of quantization levels can produce better results than directly quantizing the actual coordinates | υη, so the effective resolution of the quantizer Do it without increasing the quantization level. Instead, you can reduce the number of quantization levels to achieve the same fidelity, and maintain the same distortion when using direct quantization ϋη. Each amount is lost to 1 Ci device The age-efficient " effective " resolution will provide a more accurate quantization error signal §. Therefore, the present invention can generate a more correct reconstruction value ϋη, which then n is finally due to the positive prediction value ϋη. More positive π 2 The best amount to go: the smallest average error. However, a -3Ten's best device, because the statistical value of en will change dramatically with the 3D model. For this problem, the quantizer is immediately followed by Reproducible compendium 46313 5 V. Description of the invention (9) Simultaneous uniform and average distortion using high-resolution approximation, as explained in A. Gersho and RM Gray's "Vector Quantization and Signal CompressiorT", has been proven The consistent quantizer achieves the minimum average distortion at a specific bit rate approximation. A demand and demand quantizer should be designed to operate at high resolution, so the number of quantization levels N is large and excessively negative. The probability of the charge will be very small, and the size of the step will be much smaller than the root mean square (rms) signal value. In fact, it is shown that the source of Rugao is not memorable, and a simple consistent quantizer with a consistent encoder follows For a specific average distortion, a bit rate of about 0.25 bit can be generated from the lowest achievable bit rate (the lower limit of Shann on). Therefore, at least for a large bit rate, for no memory Source This variable rate design including a consistent quantizer with a consistent encoder behind is almost optimized. 0 Assuming the input is static and is processed by each state {Xn}, the quantized The theory {Q (χπ)} is also static and various states. Assume that HQ is the average amount processed by the quantizer, h (XQ) is the average amount of the input and D is the average distortion. Use "Vector Quantization" and Signal Compression "

Hq /? (Xq) — — l0gl2Z) where the above equality can be achieved if and only if the quantizer is consistent. Therefore, a simple uniform quantizer was selected as the quantization design. However, the dynamic range of the uniform quantizer must still be selected. If selected

Page 13 4.6 31 3 5 V. Explanation of the invention (ίο) --- The dynamic range of the quantizer is too close to match the dynamic range of the input value, which can achieve low distortion but sacrifice high bit rate, according to equation (丨). Furthermore, it is difficult to estimate the dynamic range of the input value without a second process. To avoid the risk of overloading the quantizer (one of the basic assumptions of a quantizer that has a consistent encoder), the dynamic range of the quantizer must be carefully selected. In particular, the dynamic range of this quantizer should be greater than the input data but not too large, so the reduction in distortion will be too small for the practical benefit of this design. To further improve the geometric encoder, the present invention proposes to use a "rough" quantizer Q. This rough quantizer Qr rough floating-point value is the closest fixed-length integer, instead of truncating this value. In the 3DMC / The quantization design Qt used in the TR geometry encoder truncates floating point values. This truncation process generates more quantization errors than rough floating point values. Improved Geometry Decoder FIG. 2B illustrates a block of the corresponding geometry decoder of the present invention. Illustration.

Referring to FIG. 2B, the quantized prediction error of the consistent encoding (compressed) is first decompressed to generate an uncompressed quantized prediction error. Next, the quantized prediction error in is dequantized to form a prediction error value ^. The adder then adds the prediction error ^ to the predicted geometric value 4 of the predictor P to produce the final reconstructed geometric value C) n. It should be noted that the geometric decoder of the present invention is actually equivalent to a part of the corresponding geometric encoder 'except that the inverse quantizer is compared by comparing the dotted line region of FIG. 2B and the dotted line region of FIG. 2A. By comparing Fig. 1B with Fig. 2B ', it can be found that with some modifications, the 3DMC / TR geometry decoder can be converted into a DPCM geometry decoder. specific

Page 14 46313 5

V. Description of the invention (11) 'You only need to change the order of the quantizer block and prediction 3DM (VTTR geometry decoder to DpcM decoder 4 and you should notice that MPEG4 3DMC bit 4 ... The 70-stream method does not need to make any changes. However, in the specific example, the 'dynamic range used' can be inserted into the bit stream, which is in the σσ °°! R r «Buban Wucheng, so decode The device can know the moving edge range of the code. For practical reasons, the target dynamic range can be set in advance to avoid any changes in the 3MC syntax. The efficiency of the encoding / decoding system was performed using several three-dimensional object models. Several experiments were performed using the 3DMC / TR geometric encoder proposed to the MPEG4 Commission and the modified 3DMC geometric encoder added to the teaching of the present invention. Compressing certain 3D object models. In particular, the modified 3MC geometry encoder uses 叩 CD technology with a rough quantizer. Test on 15 different 3D models. Fourteen of those 3D models MPEG4 is used in the core experiments of 3DMC. These three-dimensional object models carry any color and orthogonal information.-The size of the three-dimensional model used in the test ranges from 100,000 triangles to 20,000 squares. For objective estimation, different quantization levels are used, which are also 1, 3, 5, 8, 10, 16, 20, 25, and 30 bits. Error signals en from most data sets It was found to range from _〇 · 〇5 to _〇, 〇5, so a dynamic quantizer with a dynamic range from -0.6 to 0.6 was used to avoid the risk of overload. Subjective estimates show that this information is very good Difficult to recognize, if the majority of the three-dimensional object model in the small group is changed to an eight-bit quantization level to quantify

O: \ 64 \ 6479l.ptd

Page 15 46 313 5 V. Statement of Invention (12). Similarly, when the quantization level is too low, the previously explained high-resolution approximation assumption is no longer valid and the optimality of the uniform quantizer is broken. When the quantization level is higher than twenty bits, the bit rate may become a bit too high for the practical benefits of compression. As a result, although a wide range of quantization levels can be used, only results with quantization levels ranging from 8 bits to 20 bits can be used in the analysis. The name of the 3D object model, the size of the number of triangles, and the bit rate reduction of DPC M and the approximate quantizer V using the maximum squared quantization error and the mean squared quantization error criterion are listed in Table 1. ° The number of 3D drawing object triangles Bit rate reduction at DPCM (maximum squared quantization error criterion) Bit rate reduction at DPCM (averaged squared quantization error criterion) Elephant 29308 7.5% -16.25% 5% -18.75% Crocodile 34404 7.5% -20% 6.7 % -20 ° / 〇 'Wom-hand 2878 8.5% -13% 7.5% -12% Porsche 10474 8% -12.5% 8% -15% Cam handle 101870 9% -20% 7.5% -20% Monster 49478 7.5% -16.7% 8% -16.70 / 〇 skull 22104 8.5% -16% 7.5% -16% femur 7798 6% -16.7% 6% -16.7% electric drill 32652 8.5% -20% 8.6% -20% Balloon 12967 8.3% -18% 9.0% _18% General 22253 6.5% -11.1% 6.5% -11.8% Sanqilong 5660 4.8% -13.3 ° / 〇4.8% -12.5% Horse 22258 7.5% -20% 7.5% -20 % 57 Chevrolet 29303 2% -5% 4% -5% 83_ Honda 13594 6% -12% 5.5% -12%

Page 1646313 5 V. Description of the invention (13 ^ ~ ~ ----- The first block of f 1 lists the names of the three-dimensional object models tested. Figures 3 A and 4A Wireframe perspective of the three-dimensional model. The second column of Table 1 lists the number of triangles of different object models. The third and fourth pastes of Table 1 list the use of the DPCM geometric encoder of the present invention instead of M-PEG4 3DMC. In the case of standard geometric encoders, the percentage reduction in bit rate ◊ The third and fourth columns of Table 1 show the approximate quantizer ^ The geometric encoder consistently outperforms the proposed 3DMC / T {? Geometry encoder 'for a specific maximum squared quantization error and mean squared quantization error β The percentage range in which the bit rate of the DP CM and the rough quantizer Qr can be reduced. The maximum squared quantization error can be used. The criterion of the mean squared quantization error is from 4% to 20%. For most data sets, the absolute reduction in bit rate using DPC M and rough quantizer Qr is almost a fixed value in the area of interest, so Higher percentage reduction corresponds to Low bit-rate regions' and lower percentage reductions correspond to higher bit-rate regions "Figures 3B and 4B illustrate the bits plotted relative to the maximum squared quantization error corresponding to the different geometric codes of the electric drill of Figure 3A / Triangular table. Similar 'Figures 3C and 4C illustrate the bit / triangular graphs plotted relative to the mean squared quantization error of the different geometric codes of the electric drilling machine of Figure 3A. From Figures 3B, 3C, 4B and 4C It is self-evident that the DPCM geometric tablet composer of the present invention consistently outperforms the 3DMC / TR system. The foregoing has explained that an effective method for encoding a three-dimensional graphics model is intended for those skilled in the art to use the materials of the elements of the present invention. And configuration changes and modifications without departing from the scope of the invention.

Page 17

Claims (1)

46313 5 VI. Scope of Patent Application 1. A method for encoding a three-dimensional graphic model, the method comprising: comparing a geometric value from the three-dimensional graphic model with a predicted geometric value; and an error between the geometric value and the predicted geometric value Difference quantization produces quantization error differences. 2. As in the method of applying for the first item of the patent scope, the method further includes: consistently encoding the quantized error difference. 3. The method according to item 1 of the patent application, wherein quantifying the error difference includes performing a rough quantization. 4. The method according to item 1 of the patent application, wherein the predicted geometric value comprises a combination of at least one previously predicted geometric value. 5. The method as described in item 1 of the patent application scope, wherein the method uses the MPEG4 3D model coding syntax. 6. —A method of decoding a three-dimensional graphics model, the method comprising: dequantizing a quantization error difference to generate an error difference between a desired geometric value and a predicted geometric value; and adding the error difference to the predicted geometric value to generate a decoding The geometric value. 7. If the method of claim 6 is applied, the method further comprises: coherently decoding the quantization error difference of the encoding compression to generate the quantization error difference. 8. A method as claimed in claim 6 wherein the predicted geometric value includes a combination of at least one previously decoded geometric value. 9. The method of claim 6 in which the method uses MPEG4 Page 18 4 6 3 13 5 VI. Scope of patent application 3 D model coding syntax. 1 0. A device for encoding a three-dimensional graphic model. The device includes: a predictor that generates a predicted geometric value; and a quantizer that compares the actual geometric value with the predicted geometric value. The error difference is quantized to produce a quantization error difference. 11. If the device of the scope of application for patent No. 10, the device further comprises: a consistency encoder ', the consistency encoder is used to compress the difference in quantization error. 12. As for the agricultural property with the scope of application for item 10, the quantizer performs a rough quantization. 1 3. The apparatus of claim 10, wherein the predictor generates the predicted geometric value from at least one previously predicted geometric value. 14. The device according to item 10 of the scope of patent application, wherein the device is implemented as MPEG4 3D model coding syntax. 15. A device for performing geometric decoding, the device includes: a dequantizer that dequantizes a quantized error difference to generate an unquantized error difference between a desired geometric value and a predicted geometric value; And an adder that adds the error difference to the predicted geometric value to generate a decoded geometric value. 16. The device according to item 15 of the patent application scope, further comprising: a coherent decoder that decodes and encodes a compressed quantization error difference to generate the quantized error difference. 1 7. The device according to item 15 of the patent application scope, further comprising: Page 19 46313 5 VI. Scope of Patent Application A predictor that generates the predicted geometric value from a combination of at least one previous predicted geometric value. 1 8 _ The device according to item 15 of the patent application scope, wherein the predicted value is a combination of at least one previously decoded geometric value. -1 9. The device according to item 15 of the scope of patent application, wherein the device implements MPEG4 3D model coding syntax. Page 20