TWI504233B - Depth estimation method and device using the same - Google Patents

Description

Depth estimation method and device thereof

The present invention relates to a depth estimation method and a device therefor, and more particularly to a depth estimation method for finding an initial depth value via an object-based depth estimation and a device using the same.

In today's fast-changing technology era, stereoscopic multimedia systems are gradually being valued by the industry. In general, in the application of stereoscopic video/video, the dual-view image matching (Stereo Matching) image processing technology is a stereoscopic image core technology that is urgently needed in the industry. In the prior art, the dual-view image comparison technique first calculates an image depth map according to the dual-view image.

In general, depth maps are critical to the quality of stereo images. Based on this, how to design a dual-view image contrast depth estimation method with low computational complexity is one of the industries' dedication.

The invention relates to a depth estimation method and a device thereof, which are used for: dividing a target frame data into a plurality of object partitions, and finding an initial depth value for each object partition by an object comparison operation; The object partition determines the same group information, which corresponds to each object partition to one or more groups of the same group; and the refinement depth is generated according to the initial depth value of each object partition and other initial depth values inside the same group of objects value. Accordingly, the depth estimation method and apparatus of the present invention have the advantage of finding the initial depth value and the reference group information to generate the refined depth value through the object comparison operation compared to the conventional depth estimation method.

According to a first aspect of the present invention, a depth estimation method is proposed for performing depth estimation on dynamic multi-view image data. The depth estimation method includes the following steps. First, the dynamic multi-view image data is decompressed to find the target frame data and the sub-view target frame data. Then, the target frame data is divided into multiple object partitions, and object segmentation information is generated. Then, it is judged whether the target frame data is a class I frame or a P class/B class frame. When the target frame data is a class I frame, the reference object segmentation information is used to perform depth estimation on the target frame data to find the depth distribution data corresponding to the target frame data. When the target frame data is a P-class/B-type frame, the depth distribution data corresponding to the target frame data is found by referring to the motion vector information corresponding to the target frame data and the depth distribution data of the previous target frame data.

According to the second aspect of the present invention, another depth estimation method is proposed for performing depth estimation on multi-view input image data, and the multi-view input image data includes target frame data and sub-view target frame data. The depth estimation method includes the following steps. The target frame data is first divided into multiple object partitions. Then, the object comparison is performed on each object partition in the sub-view target frame data to correspond to the object partitions in the object partition, and the initial depth values of the successfully succeeded object partitions are determined. Then, for the object partition that fails the comparison, the initial depth value of each object partition that fails the comparison is determined by referring to the initial depth values of the plurality of adjacent object partitions adjacent to the object partition that failed the comparison. The same group information is then determined for each object partition, which is used to partition each object into at least one group of objects. Then, for each object partition, a refined depth value is generated according to the initial depth value of each object partition and other initial depth values in the same group of objects.

According to the third and fourth aspects of the present invention, there is provided a depth estimating apparatus for performing the depth estimating method according to the first and second aspects of the present invention.

In order to better understand the above and other aspects of the present invention, the preferred embodiments are described below, and in conjunction with the drawings, the detailed description is as follows:

Please refer to FIG. 1 , which is a block diagram of a depth estimating apparatus according to an embodiment of the present invention. The depth estimating device 1 includes, for example, a processor 10, a memory 20, and a bus path 30. For example, the depth estimation method in this embodiment can be implemented in a coded manner and stored in the memory 20; the processor 10 accesses the code in the memory 20 via the bus path 30, thereby implementing the implementation. Example of depth estimation method.

Next, a series of different embodiments will be further described to further describe the depth estimation method performed by the depth estimating device 1 of the present embodiment.

First embodiment

The depth estimation method and apparatus of this embodiment are used for depth estimation of static input image data. For example, the input image data is dual-view image data, including target frame data F1 and corresponding secondary view target frame data F2.

Referring to FIG. 2, a flow chart of a depth estimation method according to a first embodiment of the present invention is shown. The depth estimation method includes the following steps. First, as step (a), the processor 10 performs an image segmentation operation to divide the target frame material F1 into a plurality of object partitions A_1, A_2, ..., A_n, where n is a natural number. For example, the processor 10 determines whether it has a similar color to other adjacent pixels by referring to the color information corresponding to each pixel in the target frame data F1, thereby determining whether it belongs to the same object partition. . In an example of operation, the schematic of the target frame data F1 can be as shown in FIG. 3, where n is equal to 9, and the object partitions A_1 and A_9.

Then, as in step (b), the processor 10 compares the objects in the sub-view target frame data F2 for each object partition A_1-A_n to find the parallax for each object partition in the object partition A_1-A_n. The value Dis is used to determine the initial depth value Di of each object partition that is successful. For example, the processor 10 applies a Normalized Cross-Correlation (NCC) to implement the alignment operation in the step (b), wherein the correlation coefficient r can be expressed, for example, by the following equation (1):

Where G _s (i, j) is the source sample information, for example, the pixel data corresponding to the pixel position (i, j) in the target frame data F1; G _t (i, j) is the target sample information, that is, The pixel data corresponding to the pixel position (i, j) in the sub-view target frame data F2; and Represents the average of the internal information of the source sample and the average of the internal information of the target.

The object comparison operation in step (b) is performed, for example, in a lateral search range, thereby corresponding to each object A_1-A_n, with reference to the correlation coefficient r to find the optimum disparity value. For example, the horizontal search range is a range in which the pixel position in the x direction is between -60 and +60; in the comparison operation of the object A_1, the processor 10 partitions the object in the target frame data F1. A_1 sequentially compares with the target object of the horizontal-translation target frame data F2 and horizontally shifts from -60 to +60 pixel coordinates, thereby correspondingly finding 121 correlation coefficient values r(-60), r( -59), ..., r(-1), r(0), r(1), ..., r(60).

The processor 10 further finds the maximum correlation coefficient value r_max among the correlation coefficient values r(-60)-r(60), and determines whether the maximum correlation coefficient value r_max is substantially greater than the threshold value; if so, the processor 10 The maximum correlation coefficient value r_max corresponds to the disparity value as the initial depth value Di of the object partition A_1. When the maximum correlation coefficient r_max is less than the threshold value, the processor 10 determines that the alignment operation of the object partition A_1 is a failure. In addition, the threshold value can be a built-in value, but can also be adjusted by the user.

Then, as in step (c), for the object partition that failed the comparison, the processor 10 determines the failure of the comparison using the initial depth value Di of one or more adjacent object partitions adjacent to the object partition that failed the comparison. The initial depth value of the object partition.

In the example of the target frame data F1 shown in FIG. 3, the object partition A_1 is, for example, an object partition that fails the comparison, and the object partition A_2-A_6 is an object partition adjacent to the object partition A_1, and the object partition A_7- A_9 is an object partition that is not adjacent to the object partition A_1. In the operation of step (c), the processor 10 finds the correlation coefficient between the object partition A_1 and the adjacent object partition A_2-A_6, and sets the initial depth value of the object partition A_1 to the adjacent object partition with the highest correlation coefficient. The initial depth value Di (for example, the object partition A_2).

Accordingly, the processor 10 can provide an initial depth value Di for each of the object partitions A_1-A_n via the operations of steps (b) and (c).

Then, as step (d), the processor 10 determines the same group information Ig_1, Ig_2, ..., Ig_n for each object partition A_1-A_n, and the same group information Ig_1-Ig_n is used to respectively correspond the object partition A_1-A_n to at least one Partition object partitioning. For example, the processor 10 refers to the color information of each object partition A_1-A_n to determine whether they belong to the same group, thereby generating the same group information Ig_1-Ig_n, wherein the color information of the object partition A_1-A_n is, for example, CIELUV. The color space represents the color information.

Since the operation of the processor 10 to generate the corresponding group information for each object partition in the target frame data F1 is substantially the same, the processor 10 finds only the object partition in the target frame data F(f). The case of the same group information Ig_1 of A_1 is taken as an example.

In step (d), the processor 10 first finds a partition adjacent to the object partition A_1 in the target frame data F1, for example, an object partition A_2-A_6. The processor 10 then refers to the color information of the object partitions A_2-A_6 and A_1, and determines whether the object partitions A_2-A_6 and the object partition A_1 belong to the same group by the following equation (2):

|L _s -L _t |+|U _s -U _t |+|V _s -V _t |<Threshold (2)

L _s , U _s , V _s are color information corresponding to the object partition A_1, which is represented by CIELUV color space coordinates; L _t , U _t , V _t are corresponding to one of the object partitions A_2-A_6 Color information, which is also represented by the CIELUV color space.

When equation (2) is established, it indicates that the object partition (e.g., object partition A_2) has a similar color to object partition A_1, and processor 10 considers it to belong to the same cluster. Through a similar operation, the processor 10 sequentially substitutes the color information corresponding to each object partition A_2-A_6 into equation (2) to find the same group object partition that belongs to the same group as the object partition A_1. For example, the object partitions A_2, A_3, and A_4 are included.

Then, as in step (e), the processor 10 adjusts the initial depth values in the same group of objects for the initial depth values of the respective object partitions in the same group of objects to generate refined depth values for each object partition. . In other words, for each object partition, the processor 10 generates a refined depth value D_ref according to the initial depth value Di of each object partition and the initial depth value corresponding to each object partition in the same group partition. For example, the object partitions A_1-A_4 belong to the same same group object partition A; and for the operation of the object partition A_1, the processor 10 generates the refined depth value D_ref according to the initial depth value of the object partitions A_1-A_4. For example, the processor 10 calculates the Sum of Absolute Differences (SAD) of the initial depth value Di of the object partition A_1 and the initial depth values of the object partitions A_2-A_45. In other words, the processor 10 finds the minimum of a plurality of (for example, three) SADs for the object partition A_1 and replaces the initial depth value Di of the object partition A_1 with its corresponding initial depth value. Further, the SAD calculation operation of the processor 10 can be expressed, for example, by the following equations (3)-(7):

SAD _{A_1} =|Di _{A_1} -Di _{A_2} |+|Di _{A_1} -Di _{A_3} |+|Di _{A_1} -Di _{A_4} | (3)

SAD _{A_2} =|Di _{A_2} -Di _{A_1} |+|Di _{A_2} -Di _{A_3} |+|Di _{A_2} -Di _{A_4} | (4)

SAD _{A_3} =|Di _{A_3} -Di _{A_1} |+|Di _{A_3} -Di _{A_2} |+|Di _{A_3} -Di _{A_4} | (5)

SAD _{A_4} =|Di _{A_4} -Di _{A_1} |+|Di _{A_4} -Di _{A_2} |+|Di _{A_4} -Di _{A_3} | (6)

SAD=Min(SAD _{A_1} ,SAD _{A_2} ,SAD _{A_3} ,SAD _{A_4} ) (7)

Accordingly, the depth estimating apparatus 1 of the present embodiment can efficiently find the corresponding refined depth value D_ref for the object partition A_1 via the operations of the foregoing steps (a)-(e). Based on the similar steps, the depth estimation apparatus 1 of the present embodiment can also find the corresponding refined depth value D_ref for the other object partitions A_2-A_n in the target frame data F1, thereby completing the depth for the target frame data F1. Estimated operation.

The depth estimation method and device of the embodiment are: dividing the target frame data F1 into a plurality of object partitions, and finding an initial depth value for each object partition through the object comparison operation; determining for each object partition The same group information, which corresponds to each object partition to one or more groups of the same group; and the refined depth value is generated according to the initial depth value of each object partition and other initial depth values of the same group of objects. Accordingly, compared with the conventional depth estimation method, the depth estimation method and apparatus of the present embodiment have the advantage that the initial depth value and the reference group information can be found through the object comparison operation to generate the refined depth value.

Second embodiment

The depth estimation method of the present embodiment is different from the method corresponding to the first embodiment in that it further includes further processing based on the pixel of the refining depth value. Please refer to FIG. 4, which is a flow chart showing a depth estimation method according to a second embodiment of the present invention. For example, the depth estimation method of this embodiment further includes steps (f) and (g) after the step (e).

In step (f), the processor 10 refers to the refinement depth value D_ref, and performs a pixel-based image depth estimation for the target frame data F1 to correspond to each of the target pixels in the target frame data F1. A single pixel depth value D_pixel. Then, as step (g), the processor 10 performs a smoothing operation on the pixel depth value D_pixel corresponding to the target frame data F1 to determine the depth distribution data D(t).

Referring to FIG. 5, a partial flow chart of a depth estimation method according to a second embodiment of the present invention is shown. Further, in step (f), for example, sub-steps (f1)-(f3) are included. First, as step (f1), the processor 10 determines the search range R by using the refined depth value D_ref, and calculates a plurality of matching cost values C for each of the pen pixels in the target frame data F1, wherein each pair is matched. The cost values correspond to a plurality of different disparity values, respectively.

Please refer to FIG. 6 , which illustrates a schematic diagram of the target frame data F1 and the secondary view target frame data F2 . For example, the pixel data P(i, j) in the target frame data F1 corresponds to the coordinate position (i, j), and correspondingly has a refined depth value D_ref(i, j), wherein the target frame The data F1 has, for example, an M×N pixel array, and i and j are natural numbers less than or equal to M and N, respectively. The pixel data P'(i,j) in the sub-view target frame data F2 also corresponds to the coordinate position (i,j), and the search range R is the coordinate position (i+d_ref(i,j),j). The center, the left and right width is the search range of r. The processor 10 compares the pixel data P(i,j) with each 2r-1 pixel data in the search range R to obtain a 2r-1 pen pairing cost value corresponding to the 2r-1 pen parallax value. C(d(i,j), -r+1), C(d(i,j), -r+2),...,C(d(i,j),0),..., C(d(i, j), r-1), wherein the operation of the processor 10 to find the value of each pen can be expressed by the following equation (8):

C(d(i,j),z)=|P(i,j)-P'(i+d-ref(i,j)+z,j)|, -r<z<r (8)

For example, the width r is equal to the refining depth value D_ref(i,j), and the processor 10 corresponds to the pixel data P(i,j) to find the 2r-1 pen pairing cost value C(d(i,j) ), -r+1) to C(d(i,j), r-1). Accordingly, based on the similar operation, the processor 10 can find the 2r-1 pen pairing cost value for each M×N pen pixel data in the target frame data F1; for example, this M×N×(2r- 1) The pen pairing cost value can be represented by the three-dimensional data structure of Fig. 7.

Next, as in step (f2), the processor 10 applies a filter to perform a filtering operation on the complex pairing cost values corresponding to the same disparity values in the pairing cost values. For example, with the M×N pen value C(d(1,1), r-1)-C(d(M,) corresponding to the pairing cost value C(d(i,j), r-1) For N), r-1), the processor 10 applies a Box Filter to the value of this M×N pen C(d(1,1), r-1)-C(d( M, N), r-1) performs a filtering operation.

Then, as step (f3), in this embodiment, the processor 10 applies a dynamic programming algorithm to find a corresponding pixel depth value D_pixel for each pixel in the target frame data F1. For example, the dynamic programming algorithm finds a path corresponding to the accumulated value of the minimum alignment pair value for the paired cost value plane corresponding to the same Y coordinate value in the three-dimensional data structure shown in FIG. Find a corresponding disparity value for each X coordinate value. In the example shown in FIG. 8, it is, for example, a paired cost value plane whose Y coordinate value is equal to 1, and the disparity values corresponding to the X coordinate values X1-X5 are, for example, d2, d3, d3, d2, respectively. D3 and d1.

According to the operation of the foregoing steps (a)-(g), the depth estimating apparatus 1 of the present embodiment can effectively find the corresponding disparity value for each pen element data, and correspondingly determine the pixel depth thereof. The value is D_pixel.

For the depth estimation method of the embodiment, the operations of steps (f) and (g) need to consume the enormous computing resources of the processor 10. Accordingly, the depth estimation method of the present embodiment further provides a judging step between steps (e) and (f), for example, to determine whether to perform a higher computing power after finding the refinement depth value D_ref. Steps (f) and (g). In one example, the determination condition of the determining step can be an instruction provided by the user.

The depth estimation method and the device of the embodiment are used to: divide the target frame data F1 into a plurality of object partitions, and find an initial depth value for each object partition through the object comparison operation; for each object The partition determines the same group information, which corresponds to each object partition to one or more groups of the same group; according to the initial depth value of each object partition and the initial depth value of each object block in the same group of objects, the refinement depth is generated. Value; and the refinement depth value is processed on a pixel basis to determine the depth distribution data. Accordingly, compared with the conventional depth estimation method, the depth estimation method and apparatus of the present embodiment have the advantage that the initial depth value and the reference group information can be found through the object comparison operation to generate the refined depth value.

Third embodiment

The depth estimation method of this embodiment differs from the first and second embodiments in that it performs depth estimation for dynamic multi-view image data.

Referring to FIG. 9, a flow chart of a dynamic estimation method according to a third embodiment of the present invention is shown. The depth estimation method of this embodiment includes the following steps. First, as step (A), the processor 10 decompresses the dynamic multi-view image data MD to find the target frame data F1(t) and the sub-view target frame data F2(t) corresponding to the time t, wherein The parameter t is an integer. For example, the processor 10 decompresses the dynamic multi-view image data MD in step (A) to obtain the motion vector information I_mv of the target frame data F1(t); the processor 10 further performs the motion vector information I_mv. The processing operation is performed to perform noise filtering, and a more ideal moving vector information I_mv is obtained.

Then, as step (B), the processor 10 divides the target frame data F1(t) into a multi-object partition A_1-A_n, and corresponds to the real estate bio-partition information I_partition, wherein the object split information I_partition is used to indicate the target frame data F1. Correspondence between the pixel data in (t) and each object partition A_1-A_n. For example, the related operation of the step (B) of the embodiment is substantially the same as the step (a) in the first embodiment, and details are not described herein again.

Then, as in step (C), the processor 10 determines whether the target frame material F1(t) is a class I frame or a class P/class B frame. In general, when the target frame material F1(t) is a class I frame, the processor 10 will decompress to obtain the complete image content. Accordingly, when the target frame data F1(t) is a class I frame, the dynamic estimation method of the embodiment performs step (D), and the processor 10 refers to the object segmentation information I_partition to the target frame data F1 (t). A depth estimation is performed to find the depth distribution data D(t) corresponding to the target frame data F1(t).

Please refer to FIG. 10, which is a partial flow chart of a depth estimation method according to a third embodiment of the present invention. For example, the step (D) of the dynamic estimation method includes, for example, sub-steps (D1)-(D5), wherein step (D1) further includes sub-steps (D11), (D12), and step (D4) further includes sub- Step (D41) - (D43). In one example, sub-steps (D11), (D12), (D2), (D3), (D4), and (D5) are performed separately from steps (b), (c), (in the second embodiment). d), (e), (f), and (g) are substantially the same operations; steps (D41)-(D43) are performed substantially the same as steps (f1)-(f3) in the second embodiment, respectively. operating.

In other words, when it is determined that the target frame material F1(t) is an I frame having complete image content, the processor 10 performs the flow steps (b)-(g) in the depth estimation method of the second embodiment. In substantially the same step, the corresponding disparity value is found for each stroke data, and the pixel depth value D_pixel is determined correspondingly. And in steps (e) and (g) of the second embodiment, in this embodiment, steps (D4) and (D5) may also be performed after being judged.

In the step (C), when the target frame data F1(t) is a P class/B class frame, the dynamic estimation method of the embodiment performs the step (E), and the processor 10 refers to the target frame. The depth vector D(t-1) of the motion vector information I_mv corresponding to the data F1(t) and the previous target frame data F1(t-1) is used to find the depth distribution corresponding to the target frame data F1(t). Information D(t).

Please refer to FIG. 11 , which is a partial flow chart of a depth estimation method according to a third embodiment of the present invention. Step (E) includes, for example, sub-steps (E1)-(E5). First, as step (E1), the processor 10 refers to the object segmentation information I_partition, and finds the i-th object partition A_i in the target frame data F1(t), wherein the object partition A_i corresponds to a plurality of current object macroblocks (Macroblock). Where i is a natural number and the starting value of i is 1. Please refer to FIG. 12, which shows a schematic diagram of the target frame data F1(t) and the previous target frame data F1(t-1). For example, the object partition A_i, for example, correspondingly includes macroblocks BK1-BK6.

Then, as step (E2), the processor 10 refers to the motion vector information I_MV to find a motion vector corresponding to the current object macroblock, wherein the motion vectors respectively correspond to the current object macroblock to the previous target frame data. Multiple movements in F1(t-1) correspond to macroblocks. In the example of Fig. 12, the macroblocks BK1-BK6 in the object partition A_i correspond to the motion vectors MV1-MV6, respectively, and the motion vectors MV1-MV6 respectively point to the previous target frame data F1(t-1). The movement in the middle corresponds to the macro block BKM1-BKM6.

Then, as in step (E3), the processor 10 refers to each of the motion vectors to find a previous depth value corresponding to each of the corresponding macroblocks. In the example of FIG. 12, the processor 10 completes the depth estimation operation for the previous target frame data F1(t-1), for example, in the previous depth estimation operation, and moves the corresponding macro block BKM1. -BKM6 corresponds, for example, to the previous depth values D_M1-D_M6, respectively.

Next, as step (E4), the processor 10 finds the depth value of the object partition A_i according to the plurality of previous depth values corresponding to the target object macroblock. In the example of Fig. 12, the processor 10 averages the previous depth values D_M1-D_M6, for example, and uses this average as the depth value of the object partition A_i.

Then, as step (E5), the processor 10 increments the parameter i by 1, and repeats steps (E1)-(E4) to perform relevant depth estimation for all object partitions in the target frame data F1(t). operating.

The depth estimation method and the device of the embodiment are used for depth estimation of dynamic multi-view image data, wherein the depth estimation method and device of the embodiment decompress the dynamic multi-view image data to find out Target frame data F1(t). The depth estimation method and device of the embodiment further determine whether the target frame data F1(t) is a class I frame or a class P/class B frame; when the target frame data F1(t) is a class I frame, It is based on the object segmentation information to deeply estimate the target frame data F1(t) to find the depth distribution data D(t) corresponding to the target frame data F1(t). When the target frame data F1(t) is a P class/B type frame, it refers to the movement vector information corresponding to the target frame data F1(t) and the depth of the previous target frame data F1(t-1). The data D(t-1) is distributed to find the depth distribution data D(t) corresponding to the target frame data F1(t). Accordingly, compared with the conventional depth estimation method, the depth estimation method and apparatus of the embodiment have the advantages of referring to the decompressed motion vector information to perform depth estimation on the dynamic multi-view image data.

In conclusion, the present invention has been disclosed in the above preferred embodiments, and is not intended to limit the present invention. A person skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, the scope of the invention is defined by the scope of the appended claims.

1. . . Depth estimation device

10. . . processor

20. . . Memory

30. . . Bus path

F1, F1(t). . . Target frame data

F2. . . Secondary perspective target frame data

A_1-A_9. . . Object partition

P(i,j), P'(i,j). . . Pixel data

BK1-BK6. . . Macro block

BKM1-BKM6. . . Move corresponding macro block

FIG. 1 is a block diagram of a depth estimating apparatus according to an embodiment of the present invention.

2 is a flow chart showing a depth estimation method according to a first embodiment of the present invention.

FIG. 3 is a schematic diagram of the target frame data F1.

4 is a flow chart showing a depth estimation method according to a second embodiment of the present invention.

FIG. 5 is a partial flow chart showing a depth estimation method according to a second embodiment of the present invention.

FIG. 6 is a schematic diagram showing the target frame data F1 and the secondary view target frame data F2.

Figure 7 is a schematic diagram showing the three-dimensional data structure of the paired cost values.

Figure 8 is a schematic diagram showing the dynamic programming algorithm.

FIG. 9 is a flow chart showing a dynamic estimation method according to a third embodiment of the present invention.

FIG. 10 is a partial flow chart showing a depth estimation method according to a third embodiment of the present invention.

11 is a partial flow chart showing a depth estimation method according to a third embodiment of the present invention.

Figure 12 is a schematic diagram showing the target frame data F1(t) and the previous target frame data F1(t-1).

(A)-(E). . . Process step

Claims (17)

A depth estimation method for performing depth estimation on a dynamic multi-view image data, the depth estimation method comprising: (a) decompressing the dynamic multi-view image data to find a target frame data; (b) dividing the target frame data into a plurality of object partitions and correspondingly generating an object segmentation information; (c) determining that the target frame data is a class I frame or a class P/B frame (d) when the target frame data is the type I frame, the target frame data is deeply estimated by referring to the object segmentation information to find a depth distribution data corresponding to the target frame data; (e) When the target frame data is the P type/B type frame, refer to one of the moving vector information corresponding to the target frame data and the depth distribution data of the previous target frame data to find the target The depth distribution data of the frame data. For example, in the depth estimation method described in claim 1, the step (e) further includes: (e1) referring to the object segmentation information, and distinguishing the plurality of object partitions in the target frame data, wherein the object partitions Corresponding to a plurality of current object macroblocks (Macroblock); (e2) refer to the motion vector information to find a plurality of motion vectors corresponding to the current object macroblocks, wherein the motion vectors are used to At present, the object macroblocks respectively correspond to a plurality of moving corresponding macroblocks in the previous target frame data; (e3) refer to each of the motion vectors to find a plurality of previous depth values, the previous depth of the plurality of pens The value system corresponds to each of the corresponding corresponding macroblocks; (e4) finding a depth value of the plurality of object partitions according to the plurality of previous depth values corresponding to the target object macroblocks. For example, in the depth estimation method described in claim 1, the step (b) refers to the color information of each pixel in the target frame data to generate the object segmentation information. The depth estimation method according to claim 1, wherein the step (d) further comprises: (d1) performing image depth estimation based on the object partitioning on the target frame data, to target the target frame Each of the object partitions in the data is estimated to have an initial depth value; (d2) determining a same group information for each of the object partitions, the same group information for partitioning each of the objects to a plurality of the same group of objects; and (d3) For each of the initial depth values of each of the object partitions in the same group of objects, the initial depth values in the same group of objects are adjusted to generate a refined depth value for each object partition. The depth estimation method according to claim 4, wherein in step (a), the dynamic multi-view image data is decompressed to find a view target frame data, wherein step (d1) The method further includes: (d11) performing an object comparison on each of the object partitions in the sub-view target frame data, to find a disparity value by using the succeeding object partition, and determining the initial depth value accordingly; And (d12) determining, for the object partition in the object partition, the initial depth value of the plurality of adjacent object partitions adjacent to the object partition that failed the comparison to determine the initial of the object partition that failed the comparison Depth value. The depth estimation method according to claim 4 or 5, wherein after the step (d3), the method further includes: (d4) referring to the refining depth value, and the target frame data and the sub-angle target image The frame data is subjected to a pixel-based image depth estimation to find a corresponding pixel depth value for each of the target pixels in the target frame data; and (d5) each stroke of the target frame data The pixel depth values corresponding to the pixels are smoothed to determine the depth distribution data. The depth estimation method according to claim 6, wherein the operation of the step (d4) further comprises: (d41) determining a search range by the refined depth value, and according to the target frame data The pen pixels calculate a pairing cost (Matching Cost) value, wherein the pairing cost values have a corresponding relationship with the complex disparity values; (d42) applying a filter to correspond to the same parallax for the pairing cost values The pairing cost value of the value is subjected to a filtering operation; and (d43) the pairing cost processed by the filter is used to derive the pixel depth value corresponding to each of the pen pixels. A depth estimating device for performing a depth estimation method according to one of claims 1-7 of the patent application. A depth estimation method for performing depth estimation on a multi-view input image data, the multi-view input image data comprising a target frame data and a view target frame data, wherein the depth estimation method comprises: (a) Dividing the target frame data into a plurality of object partitions; (b) performing an object comparison on each of the object partitions in the sub-angle target frame data, so as to compare the phases in the target image frame data of the sub-view Corresponding to each object partition, and according to the comparison of the success of each object partition to generate one of the initial depth value; (c) for the object partition in the object partition failure, refer to its adjacent comparison success The initial depth value of the object partition determines the initial depth value of each object partition that fails to be determined; (d) determining the same group information for each of the object partitions, the same group information is used to map each of the objects to the plural The same group of objects; and (e) for each of the initial depth values of each of the object partitions in the same group of objects, each of the initial depths in the same group of objects Adjust the value, for each object to produce a partition refining depth value. The depth estimation method according to claim 9 , wherein after the step (e), the method further comprises: (f) referring to the refining depth value, and performing the target frame data and the sub-angle target frame data. a pixel-based image depth estimation to find a pixel depth value for each of the target pixels in the target frame data; and (g) smoothing the pixel depth values corresponding to the target frame data Operation to determine the depth distribution of one of the target frame data. The depth estimation method according to claim 10, wherein the operation of the step (f) further comprises: (f1) determining, by the refining depth value, a search range of the sub-view target frame data, and according to Each of the pen pixels in the target frame data calculates a pairing cost (Matching Cost) value, wherein the paired cost values have a corresponding relationship with a plurality of disparity values; (f2) applying a filter to target the And a pairing cost value corresponding to the same disparity value in the pairing cost value is subjected to a filtering operation; and (f3) using the filter-processed pairing cost value to obtain the pixel depth value corresponding to each of the pen pixels. The depth estimation method according to claim 9 , wherein the step (g) further comprises: referring to the object partition information to smooth the pixel depth values corresponding to the target frame data. To determine the depth distribution of one of the target frame data. The depth estimation method according to claim 9, wherein the step (a) refers to the color information of each of the pixels in the target frame data to generate the object partitions. The depth estimation method according to claim 9 , wherein in the step (b), the object comparison is performed on each of the objects in the sub-view target frame data to compare the successes. Each object partition finds a disparity value and determines the initial depth value of each object partition that is successfully aligned. The depth estimation method according to claim 9, wherein in the step (c), the object partition of the object failure in the partition of the object is referenced, and the reference is adjacent to the object partition that fails the comparison. The initial depth value of a neighboring object partition determines the initial depth value of each object partition that failed. The depth estimation method according to claim 9 , wherein in the step (d), the method further comprises: referencing one or all of the color, brightness, texture, and edge information of each of the object partitions to determine the same Group information. A depth estimating device for performing a depth estimation method according to one of claims 9-16 of the patent application.