[go: up one dir, main page]

WO2009002031A2 - Systeme de reseau systolique rapide a base de propagation des croyances, et procede de traitement de messages utilisant ce systeme - Google Patents

Systeme de reseau systolique rapide a base de propagation des croyances, et procede de traitement de messages utilisant ce systeme Download PDF

Info

Publication number
WO2009002031A2
WO2009002031A2 PCT/KR2008/003277 KR2008003277W WO2009002031A2 WO 2009002031 A2 WO2009002031 A2 WO 2009002031A2 KR 2008003277 W KR2008003277 W KR 2008003277W WO 2009002031 A2 WO2009002031 A2 WO 2009002031A2
Authority
WO
WIPO (PCT)
Prior art keywords
messages
layer
nodes
current process
buffer
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/KR2008/003277
Other languages
English (en)
Other versions
WO2009002031A9 (fr
Inventor
Hong Jeong
Sung Chan Park
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
POSTECH Academy Industry Foundation
Original Assignee
POSTECH Academy Industry Foundation
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by POSTECH Academy Industry Foundation filed Critical POSTECH Academy Industry Foundation
Publication of WO2009002031A2 publication Critical patent/WO2009002031A2/fr
Publication of WO2009002031A9 publication Critical patent/WO2009002031A9/fr
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F15/00—Digital computers in general; Data processing equipment in general
    • G06F15/16—Combinations of two or more digital computers each having at least an arithmetic unit, a program unit and a register, e.g. for a simultaneous processing of several programs
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/29—Graphical models, e.g. Bayesian networks

Definitions

  • the present invention relates to a technology for processing messages on an MRF (Markov Random Field) network having a regular array structure; and, more particularly, to a BP (Belief Propagation) based fast systolic array system and a message processing method using the same.
  • MRF Markov Random Field
  • BP Belief Propagation
  • BP is being widely used in communications systems, image processing, and the like.
  • a system is modeled as a probability network, a hidden state to be estimated is assigned to each node, and the relationship between nodes is set by a probability model .
  • MAP Maximum A Posteriori
  • a node is allocated to each pixel; and a motion vector, a disparity, a segmentation label, or the like needs to be estimated for each node.
  • a motion vector, a disparity, a segmentation label, or the like needs to be estimated for each node.
  • Such large number of nodes causes a low error rate, but too much processing time is required.
  • a 2-dimensional (2D) MRF network is defined.
  • x [x 0 Xi] ⁇ using elements X 0 and Xi
  • a position of a node is represented by a 2D vector
  • the BP is broadly classified into a max-product algorithm that estimates the MAP state and a sum-product algorithm that computes a marginal probability.
  • a description will be given for the max-product algorithm.
  • a message computation process is as in Equation 2 .
  • the size of the data cost memory becomes NiN 0 SB bits. Accordingly, the total memory size becomes 5NiN 0 SB bits.
  • the conventional BP technique shows a low error rate, it requires a large memory and a lot of computation time to estimate the hidden states on the MRF network. Further, if a plurality of parallel processors accesses the large memory to increase a processing speed, the bandwidth of the data bus is remarkably increased.
  • the present invention provides a BP based fast systolic array system and a message processing method using the same, in which a message processing sequence is to scan an MRF network in a direction, instead of a conventional iterative manner, so that a memory size can be reduced and a message can be processed by using a compact distributed memory in a VLSI (Very Large Scale Integration) chip.
  • VLSI Very Large Scale Integration
  • a belief propagation based fast systolic array system wherein a Markov random field network is constructed as a dynamic Bayesian network in consideration of an iteration axis, and messages on the dynamic Bayesian network are updated while the Markov random field network is scanned in a specific axis direction.
  • a message processing method using a belief propagation based fast systolic array system including: constructing a Markov random field network as a dynamic
  • Bayesian network in consideration of an iteration axis; and scanning the Markov random field network in a specific axis direction to update messages on the dynamic Bayesian network .
  • an MRF network is constructed as a dynamic Bayesian network in consideration of an iteration axis, and messages on the dynamic Bayesian network are updated while scanning the MRF network in a specific axis direction.
  • the memory size can be reduced, so that restriction on parallel implementation within existing VLSI chips can be overcome and fast processing can be performed with a compact distributed memory in the VLSI chips.
  • the system of the present invention can be manufactured as a compact parallel VLSI chip having a small memory, such as an FPGA
  • a complex image processing system can be manufactured as an inexpensive and small device which performs fast real-time processing.
  • the present invention can be applied to the BP application having a small number of iterations, e.g., a fast converging segmentation method disclosed in Noam Shental, Assaf Zomet, Tomer Hertz, and Yair Weiss, "Pairwise Clustering and Graphical Models” , “Advances in Neural Information Processing Systems 16", MIT Press, 2004.
  • a fast converging segmentation method disclosed in Noam Shental, Assaf Zomet, Tomer Hertz, and Yair Weiss, "Pairwise Clustering and Graphical Models” , “Advances in Neural Information Processing Systems 16", MIT Press, 2004.
  • GBP Generalized Belief Propagation
  • the present invention can be widely used as an inexpensive and low-power compact system in various signal processing fields, such as sound processing or image processing, which can be modeled on a 2D MRF network.
  • Fig. 1 illustrates an MRF network having a regular array structure and a conventional BP update rule
  • Fig. 2 illustrates a layer structure in which a layer corresponds to an iteration of a message at each node shown in Fig. 1;
  • Figs. 3A to 3C respectively illustrate an explanatory view of a layer-transformed FBP (Fast Belief Propagation) structure of the structure shown in Fig. 2 and a message update sequence;
  • FBP Fast Belief Propagation
  • Fig. 4 illustrates an explanatory view of the message update sequence shown in Figs. 3A to 3C viewed from a different angle ;
  • Fig. 5 illustrates a block diagram of a BP based fast systolic array system in accordance with the present invention
  • Fig. 6 illustrates a block diagram of the internal configuration of the FBP module shown in Fig. 5;
  • Fig. 7 illustrates a flow chart of a parallel computation process for message update within a group in accordance with the present invention
  • Fig. 8 illustrates an explanatory view of the FBP sequence shown in Fig . 7 ;
  • Fig. 9 illustrates an exemplary view of an FBP parallel matching system in accordance with the present invention
  • Fig. 10 illustrates an explanatory view of the internal structure and process of the processor shown in Fig. 9;
  • Fig. 11 illustrates a flow chart of a sequential computation process for message update within a group in accordance with the present invention
  • Figs . 12A and 12B illustrate the message update processes in the structure shown in Fig. 2 and in the layer-transformed structure shown in Figs. 3A to 3C, respectively;
  • Fig. 13 illustrates a data cost reading process in the layer-transformed structure in accordance with the present invention.
  • Fig. 2 illustrates a layer structure in which a layer corresponds to an iteration of a message at each node shown in Fig. 1.
  • Figs. 3A to 3C respectively illustrate an explanatory view of a layer-transformed FBP (Fast Belief Propagation) structure of the structure shown in Fig. 2 and a message update sequence .
  • FBP Fast Belief Propagation
  • the MRF network structure shown in Fig. 1 is constructed as a dynamic Bayesian network in consideration of an iteration axis. That is, as shown in Fig. 2, the MRP network structure shown in Fig. 1 is constructed as a dynamic Bayesian network in which a layer is stacked each time message is repeatedly computed at each node, the number of iterations ⁇ t' corresponding to a layer index ⁇ l' .
  • p(l) represents the coordinate of a node at an 1-th iteration layer
  • Figs. 3A to 3C show the vertical-rearrangement result of pod) nodes according to Equation 4.
  • the relationship between a node p (1) and a node p (1-1) at a previous iteration layer which correspond to the same node on the MRF network is as in Equation 4.
  • the node p(l) and the node p(l-l) differ from each other by an offset -[I 0] ⁇ on the layer structure .
  • nodes are grouped, and nodes at the same layer in a group are then parallel-processed to obtain messages thereof.
  • messages of nodes at the previous layer in a group are read from a local buffer of the group, and messages of nodes within the adjacent group are read from a layer buffer in which messages of nodes in the previously processed group are stored.
  • messages of nodes at each layer in a previously processed group are stored in the layer buffer.
  • messages stored in the layer buffer are updated as the previous group 300 is right-shifted, i.e., as the previous group 300 moves in a positive direction of the p 0 axis.
  • messages of nodes in a current process group 310 are computed in parallel and stored in the local buffer.
  • the messages stored in the local buffer are used to process nodes at the next (upper) layer in the current process group. Further, the messages stored in the local buffer are to be stored in the layer buffer for use in processing messages of nodes in the next group .
  • the messages of the nodes at the second layer (current layer) in the current process group 310 are computed by using, according to the conventional BP algorithm, the messages of the nodes having the node indexes 2, 3, and 4 at the previous layer (the first layer) stored in the local buffer and the messages stored in the layer buffer.
  • the local buffer is updated by the computed messages of the nodes having the node indexes 2, 3, and 4 and a layer index 1, and then used in processing the top layer (the third layer) .
  • the messages stored in the local buffer are stored in the layer buffer for use in processing the next group.
  • the messages of the nodes, at the top layer, i.e. , the third layer, in the current process group 310 have been processed as shown in Fig 3B, the messages of nodes having node indexes 3 and 4 and layer indexes 0, 1, and 2, i.e., nodes in a region 310a adjacent to a next group 320, are stored in the layer buffer, as shown in Fig. 3C.
  • messages of nodes at each layer in the next group 320 are computed using the layer buffer and the local buffer reset before the start of processing the next group 320.
  • multiple nodes at a layer in a group are processed in parallel by executing parallel processors simultaneously, and, after the completion of processing the nodes at the layer, nodes at another layer (upper layer) in the group are processed.
  • the processors for processing each layer within a group are executed in parallel, and the group is sequentially scanned.
  • the messages for each layer in a region adjacent to the next group are stored in the layer buffer
  • processing is performed using the small layer buffer and local buffer while obtaining the same result as that of the conventional BP algorithm.
  • Fig. 5 illustrates a block diagram of a parallel matching system, which is an FBP scanning system, in accordance with the present invention.
  • This system includes: a data cost module 500 that receives an input image or sound signal and computes data costs on the MRF network; and an FBP module 510 that performs an FBP scanning sequence using the data costs computed by the data cost module 500 and outputs the messages and MRF hidden states fast and in parallel.
  • the data cost module 500 receives different input signals according to the applications, and computes the data costs on the basis of the received signals. For example, in case of stereo vision, left and right images are input, and differences between absolute pixel values corresponding to each other for each disparity level are computed as the data costs . In case of motion estimation, frame images at different timings are input, and differences between absolute pixel values corresponding to each other for each motion vector state are computed as the data costs. Further, in case of sound recognition, differences between different phonemes are computed as the data costs .
  • the FBP module 510 that performs the FBP scanning sequence in the above-described manner includes: a layer buffer 600 that stores messages and data costs of nodes within a region in a previous group, the region adjacent to a current process group; a local buffer 610 that stores messages and data costs of nodes at a previous layer within the current process group, wherein the previous layer is a layer stacked directly below a currently processed layer (hereinafter, referred to as "current layer"); a group initialization unit 620 that initializes a group before the start of message update for nodes within the group,- a message update unit 630 that updates messages for each layer within a group; and a state decision unit 640 that receives the messages from the message update unit 630 and estimates hidden states at a final layer.
  • a layer buffer 600 that stores messages and data costs of nodes within a region in a previous group, the region adjacent to a current process group
  • a local buffer 610 that stores messages and data costs of nodes at a previous layer within the current process group
  • the message update unit 630 receives the data costs and the messages from the layer buffer 600 and the local buffer 610 to perform message update for the nodes at the current layer within the current process group on the basis of the received data costs and messages.
  • the message update unit 630 includes: a message computation unit 632 that receives the data costs and the messages of the previous layer within the current process group and the previous group from the local buffer 610 and the layer buffer 600, computes messages of the current layer within the current process group, and stores the computed messages in the local buffer 610; and a buffer update unit 634 that stores in the layer buffer 600 the messages of the previous layer stored in the local buffer 610.
  • the data costs and the messages of the previous layer received by the message computation unit 632 are messages of the previous layer in the current process group and in the previous group in the layer-transformed structure.
  • the message computation unit 632 accesses the local buffer 610 to read the messages or the data costs of the previous layer in the current process group and accesses the layer buffer 600 to read the messages or the data costs of the previous group .
  • the buffer update unit 634 stores, in the layer buffer 600, the messages and the data costs of the nodes within a region, in the current process group, adjacent to a next group, thereby updating the layer buffer 600.
  • the state decision unit 640 decides the hidden states of the final layer by using the messages and the data costs stored in the layer buffer 600 and the local buffer 610.
  • the entire MRF network is constructed as a dynamic Bayesian network, and the dynamic Bayesian network is divided into a plurality of groups, that is, the number of groups G 0 is determined (step S700) .
  • message update is performed on messages of nodes within each group. This process is as follows.
  • a group index g 0 is set to zero (step S702) and a layer index 1 is set to one (step S704) .
  • the group initialization unit 620 initializes a first group by reading the data costs from the data cost module 500 and setting the messages to a preset value, e.g., "0" (step S706) .
  • the message computation unit 632 updates the messages of the nodes at a first layer within the first group
  • step S708 the message computation unit 632 determines whether or not the layer index 1 is smaller than the number of layers L, i.e., whether or not the message update has been completed for all layers within the first group (step S710) .
  • step S712 the layer index 1 is increased by one (step S712) , and the message update is performed on the nodes at a second layer. That is, the steps
  • the state decision unit 640 executes a state estimation function to estimate the MAP hidden states at the final layer (step S714) .
  • the state decision unit 640 determines whether or not one or more groups for which MAP hidden state estimation needs to be performed remains , i.e., whether or not the group index g 0 is smaller than the number of groups G 0 (step S716) .
  • step S716 If it is determined in the step S716 that one or more groups for which MAP hidden state estimation needs to be performed remains, i.e., the group index g 0 is smaller than the number of groups G 0 , the group index g 0 is increased by one (step S718) , and the message update is performed on the nodes within a second group and the MAP hidden state estimation is performed for the second group. That is, the steps S704 to S716 are repeatedly performed until the group index g 0 reaches the number of groups G 0 .
  • the dynamic Bayesian network is divided into a plurality of groups, and iterations are performed within a group .
  • the edge cost Vhs (clh, d.s) / the data cost Dh (dh) , and neighboring messages trihs (d s ) are needed in order to compute the MAP state d p ⁇ L) and the message ⁇ s (d s ).
  • the edge cost V hS (d h ,d a ) is read according to the above-described FBP scanning sequence.
  • the edge cost V hS (d h ,d s ) is a constant parameter function that is not affected by the position of the node, the edge cost may be computed in advance using a given parameter.
  • the edge cost V hs (d h ,d s ) does not need to be read according to the FBP scanning sequence. Therefore, in case of, e.g. , motion estimation or stereo matching where the edge cost V hS (d h ,d s ) is computed by a truncated linear function using parameters ⁇ v and K v , additional memories are not needed.
  • the message computation and group initialization performed by the group initialization unit 620, the message computation unit 632, and the buffer update unit 634 in the FBP module 510 will be described below.
  • a node on the MRF network corresponds to nodes having an offset - [1 0] ⁇ between different iteration layers in the layer-transformed structure, as shown in Equation 5. Accordingly, N b (h)/s is changed to N b (h- [10] ⁇ ) / (s- [10] ⁇ ) by the layer transformation. Access of the message computation unit 632 to the layer buffer 600 and the local buffer 610 will be described below.
  • U 0 of a node is in a range -2 ⁇ U 0 ⁇ H 0 . If a node is within a group, i.e., 0 ⁇ u 0 ⁇ H 0 , the data cost and the message of the node at the previous layer are read from the local buffer 610. If the node is out of the group, i.e., -2 ⁇ U 0 ⁇ 0, the data cost and the message of the node at the previous layer are read from the layer buffer 600.
  • the data cost D h (d) may be read from the previous layer D h _. o ⁇ (d) (see Equation 4) . That is, similarly to the message, if 0 ⁇ h o -l ⁇ H 0 , the data cost is read from the local buffer 610. Meanwhile, if h o -l ⁇ 0, the data cost is read from the layer buffer 600.
  • the messages computed by the message computation unit 632 are stored in the local buffer 610.
  • the update of the layer buffer 600 performed by the buffer update unit 634 will be described below.
  • messages computed in the current process group and required for processing the next group are read from the layer buffer 600. That is, messages satisfying
  • H 0 -2 ⁇ U 0 ⁇ H 0 are read from the layer buffer 600 using the index satisfying -2 ⁇ U 0 ⁇ 0 during the next group processing.
  • the messages computed during the current process group processing need to be stored in the layer buffer 600 for the next group processing.
  • the data costs satisfying H 0 -I ⁇ h 0 ⁇ H 0 are stored in the layer buffer 600 using an index satisfying -1 ⁇ ho ⁇ 0.
  • the group initialization, message computation, and message update can be represented by following functions .
  • the FBP module 510 includes a plurality of local buffers, a plurality of layer buffers, and a plurality of systolic array processors PE which access the buffers .
  • the FBP scanning sequence is performed by the FBP module 510 having a parallel VLSI architecture, as shown in Fig. 10.
  • the systolic array processors PE read the messages from the local buffers and the layer buffers of neighboring processors to perform parallel computation.
  • each of the systolic array processors PE includes the message update unit 630 and the state decision unit 640.
  • the messages of nodes satisfying -2 ⁇ u ⁇ 0 are read from the layer buffer.
  • the size of the layer buffer for all of the messages is 4NiLSB bits. Since the local buffer only stores messages in all directions of the current layer, the size thereof is 4NiH 0 SB bits. Accordingly, the message memory size is 4Ni (H o +L) SB bits .
  • the layer buffer is N x LSB bits
  • the local buffer is NiH 0 SB bits.
  • the size of the data cost memory is Ni(H 0 +L)SB bits.
  • the total memory size according to the present invention is 5Ni(H 0 +L)SB bits.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Hardware Design (AREA)
  • Software Systems (AREA)
  • Image Processing (AREA)

Abstract

La présente invention concerne un système de réseau systolique rapide à base de propagation des croyances dans lequel on construit un réseau de Markov à champs aléatoires sous forme d'un réseau de Bayes dynamique tenant compte d'un axe d'itérations. En outre, des messages concernant le réseau de Bayes dynamique sont mis à jour pendant qu'on parcourt systématiquement le réseau de Markov à champs aléatoires dans le sens spécifique d'un axe.
PCT/KR2008/003277 2007-06-28 2008-06-12 Systeme de reseau systolique rapide a base de propagation des croyances, et procede de traitement de messages utilisant ce systeme Ceased WO2009002031A2 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
KR1020070064351A KR100920229B1 (ko) 2007-06-28 2007-06-28 Ｂｐ의 고속 시스톨릭 어레이 시스템과 이를 이용한 메시지처리 방법
KR10-2007-0064351 2007-06-28

Publications (2)

Publication Number Publication Date
WO2009002031A2 true WO2009002031A2 (fr) 2008-12-31
WO2009002031A9 WO2009002031A9 (fr) 2009-02-26

Family

ID=40186141

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/KR2008/003277 Ceased WO2009002031A2 (fr) 2007-06-28 2008-06-12 Systeme de reseau systolique rapide a base de propagation des croyances, et procede de traitement de messages utilisant ce systeme

Country Status (2)

Country Link
KR (1) KR100920229B1 (fr)
WO (1) WO2009002031A2 (fr)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104156964B (zh) * 2014-08-14 2017-03-08 陈荣元 一种综合mrf和贝叶斯网络的遥感影像区域分割方法
CN106780120A (zh) * 2016-12-06 2017-05-31 广东电网有限责任公司电网规划研究中心 输变电工程造价指数处理方法和装置
CN108269273A (zh) * 2018-02-12 2018-07-10 福州大学 一种全景纵向漫游中极线匹配的置信传播方法

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20110011463A (ko) 2009-07-28 2011-02-08 엘지이노텍 주식회사 라이트 유닛 및 이를 구비한 표시장치

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR100232096B1 (ko) * 1996-09-13 1999-12-01 박래홍 이산 웨이브렛 변환을 위한 vlsi의 구조
KR100374784B1 (ko) 2000-07-19 2003-03-04 학교법인 포항공과대학교 실시간 입체 영상 정합 시스템

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104156964B (zh) * 2014-08-14 2017-03-08 陈荣元 一种综合mrf和贝叶斯网络的遥感影像区域分割方法
CN106780120A (zh) * 2016-12-06 2017-05-31 广东电网有限责任公司电网规划研究中心 输变电工程造价指数处理方法和装置
CN108269273A (zh) * 2018-02-12 2018-07-10 福州大学 一种全景纵向漫游中极线匹配的置信传播方法
CN108269273B (zh) * 2018-02-12 2021-07-27 福州大学 一种全景纵向漫游中极线匹配的置信传播方法

Also Published As

Publication number Publication date
WO2009002031A9 (fr) 2009-02-26
KR100920229B1 (ko) 2009-10-05
KR20090000347A (ko) 2009-01-07

Similar Documents

Publication Publication Date Title
US12106482B2 (en) Learning-based active surface model for medical image segmentation
CN110837811B (zh) 语义分割网络结构的生成方法、装置、设备及存储介质
CN113826119B (zh) 纯注意力的计算机视觉
CN110991513B (zh) 一种具有类人连续学习能力的图像目标识别系统及方法
CN110874563A (zh) 通过cnn的多个图像输出的提供集成的特征图的方法及装置
CN114299303A (zh) 一种船舶目标检测方法、终端设备及存储介质
Kobayashi Noise robust projection rule for hyperbolic Hopfield neural networks
CN108764244B (zh) 基于卷积神经网络和条件随机场的潜在目标区域检测方法
CN113298097B (zh) 基于卷积神经网络的特征点提取方法、设备及存储介质
CN116109678B (zh) 基于上下文自注意力学习深度网络的目标跟踪方法与系统
CN114926636A (zh) 一种点云语义分割方法、装置、设备及存储介质
CN114119749A (zh) 一种基于密集关联的单目3d车辆检测方法
CN113449612A (zh) 一种基于子流型稀疏卷积的三维目标点云识别的方法
CN111339869A (zh) 一种人脸识别方法、装置、计算机可读存储介质及设备
WO2009002031A2 (fr) Systeme de reseau systolique rapide a base de propagation des croyances, et procede de traitement de messages utilisant ce systeme
WO2024001283A1 (fr) Procédé d'apprentissage pour réseau de traitement d'image, et procédé et appareil de traitement d'image
CN113971734B (zh) 目标对象检测方法、装置、电子设备及存储介质
CN117830703A (zh) 基于多尺度特征融合的图像的识别方法、计算机装置和计算机可读存储介质
WO2009014314A1 (fr) Système de réseau systolique rapide à base de propagation des croyances, et procédé correspondant
CN115908874A (zh) 一种基于孪生网络的目标跟踪模型去冗余方法
CN114693919A (zh) 一种目标检测方法、终端设备及存储介质
CN116993828B (zh) 基于激光雷达点云聚类的点云匹配定位方法、系统及产品
CN116805371B (zh) 激光点云和视觉图像的跨模态场景匹配方法、装置及设备
CN114677380B (zh) 一种基于多样化交互的视频对象分割方法及系统
CN115908813B (zh) 一种道路场景下的实时性语义分割系统及方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 08766239

Country of ref document: EP

Kind code of ref document: A2

DPE1 Request for preliminary examination filed after expiration of 19th month from priority date (pct application filed from 20040101)
NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 08766239

Country of ref document: EP

Kind code of ref document: A2