AI solution in the field of Computer vision

 Object detection from image and video is very challenging task today in computer vison. That is used for generally detect the object like car, person, mobile, road. Recently that is applied on autonomous car and many more. Detection and locating the object of certain class via deep learning concept improve the performance of detector .in this Blog we focus on various detection method that already exist. we doing this review from basic to advance and we go deeper via exploring various types of networksthat is used in object detection.

object detection means how to find accurate object and specific object, and determine the object at various angle and locate the correct object, real time object, progressive object, hidden object, multi object, moving object and how can we fastley detect the object ,3d object detection.so this research paper going through all this technique and give the right analysis through review of some work, so we can easily track our object.

For locating the object, we review various technique like region based, YOLO, various type convolution neural architecture R-CNN.alexnet, fast RCNN and so on. goal of this blogis to review the object detection technique that is held recently.

in object detection we focus on object in image, multipleobjects of same and different class of object in image. different types of scaled object in image.

Object detection is through various technique like sliding window bounding box technique that comes with some draw back after that SIFT and HOG method introduce for detecting the reliable object. Object detection focus on salient and generic object detection. How can detected the object using bounding box and CNN and without bounding box method for detecting the object using weighted hausedraf distance but that also has some draw back like pixel losses. We can also track the method that not depend on bounding box that is PCA sliding window-based method

 

In this blog we include the last five-year work progress in object detection field using deep learning concept on popular dataset. We provide the proper framework so we can easily track the object detection .and we also discuss some problem that is faced by this method and also discuss the Loss function thatincurred by various method that is used for object detection.

Object detection technique region proposal that is used by selective search or edge boxes by Feed ForwardArtificial neural network or some small convolution neural network or some with extra layer of CNN for locating the object. Due to some limitation of this Blog, we are not included some research work. we mainly focus on some standard and futuristic paper that will be extended by the researcher in the future after reading this review work.

Object can vary in shape and size and may be belong to different class in single image so first we focus on position then shape then size .so we use sliding window concept but due large no. of window in the image creates overbudget of detection problem so we use SIFT ang HOG features to. if detector produces high accuracy but faces ambiguity problem of classes for static and dynamic data. Also faces a problem for different intensity, low resolution object, large scale image, small image, and one more important criterion for selection is time constraint. That may be solved by Gpu

Firstly, we go through some object detection deep learning method review. From 2013 to 2020

Fig1

 

So we focus from 2015-2019 so we exacley define the research work done bu varoius researcher and we conclude our work at end by comapring on varous dataset and we explare the mAP parameter that used by this network.

 

 

 

 

 

 

Humans have a glance at an image and can differentiate between the objects present in a single frame. It is very fast and accurate. To meet to accuracy and speed of human’s detection; many algorithms have come into existence which is mainly categorized as two stage object detectors such as R-CNN family and one stage object detectors such as YOLO, SDD etc. These algorithms perform differently (accuracy, mAp changes) with different datasets and network architectures.

                Many datasets have come into picture over many years. Here we are dealing with the custom dataset with the size of 1000 having 20 class labels. The mAp of the YOLO on this dataset is slightly higher compared to its mAp value on other datasets such as COCO, Pascal VOC. Here we use Darknet architecture to detect the objects. YOLO v1, YOLO v2 (YOLO 9000), YOLO v3 are implemented on the custom dataset with different Darknet architectures giving an output of different and improvised mAp.

 

.

Fig.2

We have observed that YOLO v1 is simple and can detect some objects and can’t find small objects if they are in a cluster. YOLO v2 is an improved version of the previous one which reduces over fitting, overcomes the problem of detecting smaller objects using fined-grained features by dividing the image into  grid cells and also detects objects with different dimensions and configurations. We have also noticed that YOLO v2 and YOLO v3 have almost the same features. But, YOLO v3 has succeeded in getting implemented in real time. It detects real-time objects accurately and quickly.

 Object detection reduces the man’s effort in many aspects. Applying it in the real-time is one of the major tasks in today’s world . Many object detection algorithms have come into existence , such as SDD, Faster  R-CNN, YOLO v4,v5 etc. Enhancing these algorithms day by day can change the way of detection

Ravi Prakash