Introduction:
Object Recognition has completely changed how machines see and interact with their environment. The above-mentioned tasks becomes important, where the understanding of a 3D scene is necessary. Robotic vision, automation on shop floor, surveillance camera, autonomous vehicles etc. are the major areas where the ability to recognize and comprehend objects is critical. The vision of successful implementation in these areas could become a reality. And, this is due to advancements in deep learning and artificial intelligence-driven object recognition algorithms.
Object Recognition is a key task of Computer Vision. It involves one or combination of more than one tasks of image classification, image localization, object detection and image segmentation. Image classification is the task in which the given image is classified to particular class. In image localization task, the location of an object in the given image is the main task. Object detection is the combination of image classification and image localization. This task finds objects in the given image and assigns a label to each bounding box. Image segmentation is the task of getting the information of an image by segmenting the image into various parts, and processing the relevant parts of the image.
Traditional Approach to Object Recognition:
Traditional computer vision techniques of object recognition were based on handcrafted features known as detectors. Earlier, feature representations were drawn considering the shape and size of an object’s attributes. This detector then slides over an image to look for the matching feature in the image to detect an object in the image. Before deep-learning era, features were mainly extracted with the use of Scale Invariant Feature Transform (SIFT) and Histogram of Oriented Gradient (HOG).
The limitations of traditional approach to Object Recognition are:
- Limited feature engineering due to handcrafted features act as detectors
- Lack of Hierarchical Representation
- Scalability Issues
- Limited Generalization
- Sensitivity to Noise
Deep Learning Approach to Object Recognition:
On the other hand, Deep Learning techniques are representation-learning techniques. Here, features, and fine details in images are represented at multiple levels, from a low abstraction level to a higher abstraction level. LeCun et al.’s effort of developing Rumelhart et al.’s method of training via backpropagation with stochastic gradient descent, the increased computational capabilities, the rebirth of Convolutional Neural Networks (CNNs) in 2012 with better classification results, triggered the challenge to develop a model that can fill the gap between image classification and object detection. R. Girshick et al. took up the challenge and came up with the breakthrough by presenting the Regions with CNN features (RCNN) in 2014.
