With recent advancements in the industry of artificial intelligence, computer vision and deep learning have gained a lot of attention. To their credit, now Object Detection applications, which were earlier considered extremely challenging, have become easier to create.
Object detection can be defined as a computer vision technique which aims to identify and locate objects on an image or a video. Computers might be able to process information way faster than humans, however, it is still difficult for computers to detect various objects on an image or video. The reason for this is that the computer interprets the majority of the outputs in the binary language only.
This article aims to briefly discuss:
- The basics of object detection
- The object detection models
- The benefits of object detection
- The challenges and solutions
Before we get to the points above, we need to understand the difference between image classification and object detection. Beginners tend to confuse these two.
Difference between Object Detection and Image Classification
Let us break down these techniques, to know the difference between them. When you look at a picture of a dog you can instantly say it’s an image of an animal i.e. tell what the image is about. This is what image classification is all about.
As long as there is only one object, image classification techniques can be used.
But if we have multiple objects, that’s when the concept of Object Detection comes into play. By building rectangular boxes around the object of interest we can help the machine recognize the object each box contains. We can also indicate the exact location of the objects using this method. It is possible for a single picture to contain many objects, so multiple bounding boxes may be shown.
Object detection applications are limitless, but they generally identify and detect the real-objects such as human beings, buildings, cars and many more. Additionally, a machine needs a lot of labeled data of different kinds of objects for it to recognize those objects in the future. This means the ML model being trained on that labeled dataset will have a better chance to make accurate predictions.
Moreover, there are several companies which offer data annotation services. You just have to choose the right one based on your requirements. This technique is widely applied in people/object tracking applications, video surveillance cameras which I will elaborate further.
Object Detection Models
Now that we are clear with the definition of Object Detection, let’s have a look at some popular Object Detection models.
R-CNN, Faster R-CNN, Mask R-CNN
The most popular object detection models belong to the family of regional based CNN models. This model has revolutionized the way the world of Object Detection used to work. In the past few years, they’ve not only become more accurate but more efficient too.
SSD and YOLO
There are a plethora of models belonging to the single shot detector family which were published in 2016. Although SSDs are faster than CNN models, their accuracy rate is much lower than that of the CNNs.
YOLO or you only look once, is quite different from region-based algorithms. Just like SDDs, yolo is faster than R-CNNs but lags behind because of low accuracy. For mobile or embedded devices, SDDs are the perfect choice.
CenterNet
In recent years, these object detection models are gaining more popularity. CentreNet follows a key point-based approach for object detection.
When compared with SSD or R-CNN approaches, this model proves to be more efficient and as well as more accurate. The only drawback of this method is slow training process.
Benefits of Object detection to Real-world
Object detection is completely inter-linked with other similar computer vision techniques such as image segmentation and image recognition that assist us to understand and analyze the scenes in videos and images. Nowadays, several real-world use cases are implemented in the market of object detection which make a tremendous impact on different industries.
Here we’ll specifically examine how object detection applications have impacted in the following areas.
Self-driving cars
The primary reason behind the success of autonomous vehicles is real-time object detection artificial intelligence based models. These systems allow us to locate, identify and track the objects around them, for the purpose of safety and efficiency.
Video Surveillance
Real-time object detection and tracking the movements of objects allow video surveillance cameras to track the record of scenes of a particular location such as an airport. This state-of-the-art technique accurately recognizes and locates several instances of a given object in the video. In real-time, as the object moves through a given scene or across the particular frame, the system stores the information with real-time tracking feeds.
Crowd Counting
For heavily populated areas such as shopping malls, airports, city squares and theme parks, this application performs unbelievably well. Generally, this object detection application proves to be helpful to large enterprises and municipalities for tracking road traffic, violation of laws and number of vehicles passing in a particular time frame.
Anomaly detection
There are several anomaly detection applications available for different industries which use object detection. For instance, in agriculture, object detection models can accurately recognize and find the potential instances of plant disease. With the help of this, farmers will get notified and they will be able to prevent their crops from such threats.
As another example, this model has been used to identify the skin infections and symptomatic lesions. Some applications are already built for skin care and acne treatment using object detection models.
Keep in mind, there are some problems encountered while creating any kind of object detection model. However, solutions are also available to limit the challenges.
Challenges and Solutions of Object detection Modelling
Dual Synchronization
The first challenge for object detection is to classify the image and position of the object, which is known as object localization. In order to address this problem, most developers often use a multi-tasking loss function to penalize both localization and misclassification errors.
- Solution: Regional based Convolutional neural networks displays one class of object detection framework that consist of region generation proposals where objects are likely to be located, followed by CNN models processing to classify and rectify the object locations. Fast-R CNN model can improve the initial results with R-CNN. As its name denotes, this Fast R-CNN model provides tremendous speed, but accuracy also improves only because the localization and object classification tasks are optimized using a multi-task loss function.
Real-time detection speed
Fast speed of object detection algorithms has always been a major problem to classify and localize the crucial objects accurately at same time to meet the real-time video processing. Over the years, several algorithms improved the test time from 0.02 frames per second to 155 fps.
- Solution: Faster R-CNN and Fast R-CNN models aim to speed up the original speed of R-CNN approach. Because R-CNN uses the selective search to produce 2000 candidate regions of interest and passes through each CNN based model individually, that may cause a heavy bottleneck since the model processing gets down. Whereas, Fast R-CNN model transmits the whole image through CNN base once and then matches the ROIs created with selective search to feature map, considering 20-fold reduction in processing time.
Multiple aspects ratios and spatial scales
For several object detection applications, items of interest may appear in huge range of aspect ratios and sizes. Researchers proved numerous methods to ensure the detection algorithms which are able to recognize different objects at different views and scales.
- Solution: Rather than selective search, faster R-CNN has been updated with a region proposal network that uses a small sliding window over the picture’s convolutional feature map to produce candidate regions of interest. Several regions of Interests can be predicted at different positions and described relative to reference anchor boxes. The size and shape of these anchor boxes are selected to span a range of aspect ratios and different scales. It lets several types of objects identify with a hope that bounding box coordinates do not need to be adjusted during the localization task.
Limited data
One of the undeniable facts to be considered is the limited amount of annotated data which becomes a hurdle to build an application. These datasets are specifically containing ground truth examples for dozens to hundreds of objects, while image classification datasets include approximately 100,000 different classes.
- Solution: Well, several image datasets are available over the internet like COCO Dataset, which was offered by Microsoft that currently leads some of the object detection annotated data. This dataset contains 300,000 segmented pictures with 80 different object categories according to the precise location labels. Each image contains an average of 7 objects and items that appear at very large scales. One of the most interesting methods to reduce the data scarcity comes with YOLO9000, the second version of YOLO. YOLO9000 deals with numerous crucial updates into YOLO, but it also aims to narrow down the dataset gap between image classification and object detection. Moreover, it trains parallel both ImageNet and COCO, an image classification dataset with tens of thousands of object classes.
Final thought
According to the sources, object detection is considered much harder than classification, specifically because of the above mentioned problems. Researchers continue to apply great efforts to mitigate these obstacles, which at times have yielded amazing results; however, significant problems still persist. Certainly all object detection models are struggling with small objects, especially those collected together with partial occlusions. Real-time detection with object classification and localization accuracy is still a notable issue and researchers often prioritize one or the other thing when making design decisions. On an optimistic note, video tracking might see some further advancements in the future in a variety of other contexts.
In this post, I tried to briefly touch on the basics of object detection techniques. I really hope you found this short article helpful.
