Find Jobs
Hire Freelancers

C homework on CELL architecture

$30-180 USD

已取消
已发布将近 14 年前

$30-180 USD

货到付款
PLEASE CONTACT ME FOR RESOURCES THAT YOU NEED - WRITTEN BETWEEN [!!!xxxx]. THe complete statement i'll give you because its a limit of characters. Viola-Jones Face Detection: Face Detection Stage? The deadline for submitting theme: Friday, May 28, 2010, at 11:55 p.m. ? With a ready trained classifier, we can move to scan an image in search of faces. ? Homework 4 aims to implement a faces detector on the Cell BE architecture. 1. Classifier The classifier to be used in this theme is that of this file [!!!haar_classifier]. This classifier is derived from that used by OpenCV ? The Classifier contains several steps (stages) to check if a sub-window of the image is a face. Each stage contains several features and each ? feature contains multiple rectangles. ? THe file's format containing the classifier? is illustrated in Fig. 1 (green lines written are comments - only in figure they don't appear in the file): ? [!!![login to view URL]] Each feature of the current stage ? (initially 0) is evaluated by calculating the sum: ? [!!![login to view URL]] ,where no_rects represents the number ? of rectangles of the feature and w[i] refers to the weight associated with the rectangle rects[i].? After evaluating the feature , if the value obtained is less than your feature's threshold, then we consider left_val of this feature,? ? otherwise it is considered right_val. For the current stage is the sum of left_vals and right_vals considered and if this sum ? exceeds the stage's threshold then the sub-image is classified as face and move to your next stage (current stage + 1) . If all stages? ? have classified the image as a face, then the sub-image is considered to be face. ? 2. Details of implementation and testing? TO BE GIVEN BY ME To implement the homework we recommend ? to start from the serial version [!!![login to view URL]], where you will find more useful structures and [login to view URL] homework will be implemented on a Cell BE architecture with 8 SPU's. Image scanning looking for faces will be made in? ? parallel on the 8 SPU's.? ? Running example: ? ./ppu ./haar_classifier ./datasets/[login to view URL] 640 311 40 ? Output Example : ? 10 183 176 0.674191 To test use this archive [!!![login to view URL]] containing the folder test.? ## Deliverables Viola-Jones Face Detection: Face Detection Stage? The deadline for submitting theme: Friday, May 28, 2010, at 11:55 p.m. With a ready trained classifier, we can move to scan an image in search of faces. ? Homework 4 aims to implement a faces detector on the Cell BE architecture. ? 1. Classifier The classifier to be used in this theme is that of this file [!!!haar_classifier]. This classifier is derived from that used by OpenCV [!!![login to view URL]]:? [login to view URL] Using the Classifier: The Classifier contains several steps (stages) to check if a sub-window of the image is a face. Each stage contains several features and each ? feature contains multiple rectangles. ? THe file's format containing the classifier ? is illustrated in Fig. 1 (green lines written are comments - only in figure they don't appear in the file): ? ? [!!![login to view URL]] ? ? Classifier is used as follows:? Each feature of the current stage ? (initially 0) is evaluated by calculating the sum: ? [!!![login to view URL]] ,where no_rects represents the number ? of rectangles of the feature and w[i] refers to the weight associated with the rectangle rects[i].? ? After evaluating the feature , if the value obtained is less than your feature's threshold, then we consider left_val of this feature,? ? otherwise it is considered right_val. For the current stage is the sum of left_vals and right_vals considered and if this sum ? exceeds the stage's threshold then the sub-image is classified as face and move to your next stage (current stage + 1) . If all stages? ? have classified the image as a face, then the sub-image is considered to be face. ? ? 2. Details of implementation and testing? Because the classifier's training ? with OpenCV Branch was done with images that were processed to reduce the influence of the light, the results given by the detector implemented in this homework will not be conclusive. The detector will need to display only? information on that sub-image passing the most stages. If there are several sub-images that pass the maximum number? of stages then it will be considered the one for which the difference between Threshold's stage and the cumulative amount(sum) of stage's features? ? (left_vals and right_vals) is minimal. If this difference is the same for sub-images that pass the maximum number? ? of the stages, then it will consider the stage with the maximum x coordinate and if these are equal again - you will then consider that which has the? ? maximum y coordinate (Cartesian coordinates). Also, in order to simplify, our program will test the sub-images of a given dimension.? Because the classifier was trained with images of 20x20, for example if we want to test the sub-images of 40x40,then the features ? will have to be scaled (coordinates and sizes will be multiplied by 2 in this example). ? ? To implement the homework we recommend ? to start from the serial version [!!![login to view URL]], where you will find more useful structures and functions. ? The homework will be implemented on a Cell BE architecture with 8 SPU's. Image scanning looking for faces will be made in? ? parallel on the 8 SPU's.? As in the serial version, the program will receive five input arguments: the file containing the classifier, the file containing the image file,? image width, image height and the size of the faces wanted. Running example: ? ? ./ppu ./haar_classifier ./datasets/[login to view URL] 640 311 40 ? ? The program will display at STDOUT a single line ? that contains four data: the maximum number of stages passed by a sub-image, ? x and y of that sub-image, the difference between stage's threshould which has not passed and the amount accumulated from the features's? ? thresholds of the stage. Output Example : ? ? 10 183 176 0.674191 ? ? ? We recommend to use software managed cache because you will have to make many DMA transfers that care must be taken. Also, be sure to copy the classifier's ? ? structures on the SPU's correctly. ? ? To test use this archive [!!![login to view URL]] containing the folder test. In this folder you will copy the code's executable written for PPU under the name ppu and then you will run the "make run" command. The test will run your program twice, both for image test / datasets / [login to view URL], ? but for different sizes of the tested sub-images as faces (40 and 80). ? Warning! THe Archive that you send will be mandatory to include a Makefile to compile the homework and a readme text file that explains ? what have you done with your PPU and SPUs (transfers, synchronization, etc.). And other details that you consider relevant. 3. Score (maximum 110 points) - 80 points - correct implementation, passing the tests (40 points 40 points T1 + T2). Running tests is a timeout 120 seconds.? The homeworks will be runned on cell-QS22-X systems. [!!! you have in the virtual machine that i'll give to you a 8-SPU CELL simulator]. - 10 points - the use of SIMD instructions in at least one place - 10 points - assessing assistant (readme, implementation) - 10 points - BONUS - use of software managed cache Warning! Topics not to pass any test will be able to get up to 40 points I NEED AT LEAST 80 points ? ? ?
项目 ID: 3415189

关于此项目

远程项目
活跃14 年前

想赚点钱吗?

在Freelancer上竞价的好处

设定您的预算和时间范围
为您的工作获得报酬
简要概述您的提案
免费注册和竞标工作

关于客户

ROMANIA的国旗
Romania
0.0
0
会员自5月 10, 2010起

客户认证

谢谢!我们已通过电子邮件向您发送了索取免费积分的链接。
发送电子邮件时出现问题。请再试一次。
已注册用户 发布工作总数
Freelancer ® is a registered Trademark of Freelancer Technology Pty Limited (ACN 142 189 759)
Copyright © 2024 Freelancer Technology Pty Limited (ACN 142 189 759)
加载预览
授予地理位置权限。
您的登录会话已过期而且您已经登出,请再次登录。