Rethinking Object Detection
Date:
Object localization/detection is one of the most crucial tasks in artificial intelligence and computer vision. The importance of object detection is that most of the vision tasks start with object localization. Real-world applications such as autonomous driving, personal/industrial robotics, person counting, object tracking, surveillance, OCR (optical character recognition) need to localize the object in the given image or video.
Object detection has a long history. The researchers have been continuously working on improving object detection techniques. Before the deep learning era, hand-crafted features — haar-like features, HOGs (histogram of gradients), and deformable part models — were used to train an object localization classifier. With the great success of deep learning in computer vision, novel deep learning-based object detection methods (features extracted from deep convolutional neural networks) have been proposed. There are various very robust and high performing object detection methods which help to boost real-world applications.
Session Outline:
In this tutorial I went through:
the history of object detection — pre-deep learning methods object detection evaluation — evaluation metrics;
benchmark datasets used for evaluation;
cascade face detection methods (MTCNN);
convolutional neural networks;
the recent (deep learning-based) detection methods such as RCNN (Region Convolutional Neural Networks), Fast RCNN, Faster RCNN, RetinaNet, YOLO (You Only Look Once), SSD (Single Shot MultiBox Detector);
two-stage detectors vs. single-stage detectors;
end-to-end object detection pipeline;
practical tips and tricks in training object detection models.
Then, we have discussed future research topics, and asked and tried to answer some questions:
How could the current object localization/detection methods be improved?
What are the missing points in state-of-the-art object detection methods and datasets?
What we should pay attention to have explainable and reliable detectors
Which evaluation metrics we need to consider to be able to assess the detection methods in the right way?
