Exploring the Most Recent Deep Learning Object Detection Algorithms and appropriate regulations

The field of deep learning-based object detection is continuously evolving, with new algorithms and techniques emerging regularly. Researchers and practitioners are pushing the boundaries of speed, accuracy, and efficiency to make object detection more accessible and applicable to various real-world scenarios. As you explore the most recent deep learning object detection algorithms mentioned in this blog, keep in mind that choosing the right algorithm depends on your specific application requirements, available resources, and performance expectations. Staying updated with the latest advancements is crucial to harness the full potential of object detection in the ever-expanding realm of computer vision.

 

This blog post consists of two parts. Part A features a list with the most recent deep learning object detection algorithms and part B features some of the latest regulations relevant to this field.

Part A: Deep Learning Object Detection Algorithms

Object Detection. This involves identifying and locating objects within an image or video frame. Object detection algorithms aim to draw bounding boxes around objects and label them with their respective classes. Object detection is a fundamental computer vision task that has made significant strides in recent years, thanks to the rapid advancements in deep learning. Detecting and locating objects within images and videos has numerous practical applications, ranging from autonomous vehicles and surveillance systems to healthcare and augmented reality. In this blog post, we’ll delve into the most recent deep learning object detection algorithms that have been making waves in the field of computer vision as well as proper regulations that set a direction in EU. In the following, we list some of the most recent and robust object detection algorithms.

a. YOLOv4 and YOLOv5.
Figure 1. source: https://viso.ai/deep-learning/yolov5-controversy/

The You Only Look Once (YOLO) series of object detection algorithms have been at the forefront of real-time object detection. YOLOv4, released in 2020, and YOLOv5, which followed in 2020, have pushed the boundaries of speed and accuracy. YOLOv5, in particular, introduced a streamlined architecture, making it even faster while maintaining competitive accuracy.

b. YOLOv8 and YOLO-NAS (As of 2021).
Figure 2. source: https://learnopencv.com/ultralytics-yolov8/

 YOLOv8 and YOLO-NAS are the newest versions in the series of YOLO. Both algorithms offer high frame per second rate while applying simultaneously instant segmentation. They are not yet widely recognized or documented versions of the YOLO algorithm.

c. EficientDet.

 It was introduced in 2019, focuses on achieving excellent trade-offs between model efficiency and accuracy. It leverages a compound scaling method to efficiently balance network depth, width, and resolution. This family of models has performed exceptionally well on various object detection benchmarks while being computationally efficient.

d. DETR (Data-efficient Transformer for Object Detection)
Figure 3. source: https://medium.com/@yanzixuan1019/detr-and-efficient-detr-614a94dfde10

DETR, introduced in 2020, represents a significant shift in object detection as it replaces the traditional anchor-based methods with a transformer-based architecture. Transformers, originally designed for natural language processing tasks, have demonstrated impressive results in object detection by directly predicting object positions from the image.

e. Sparse R-CNN.

 Sparse R-CNN, introduced in 2021, is designed to reduce the computational cost of traditional Region-Based Convolutional Neural Networks (R-CNN). It leverages sparsity-inducing operations to significantly reduce the number of computations required for object detection, making it more efficient for deployment in resource-constrained environments.

f. CenterNet

CenterNet is an object detection algorithm that focuses on predicting object centers and their corresponding bounding boxes. This approach simplifies the object detection pipeline and has demonstrated excellent results in terms of speed and accuracy.

g. BlazePose

 While primarily designed for pose estimation, BlazePose can also be used for object detection tasks, especially in real-time applications. It offers both upper body and full-body pose estimation, making it versatile for various computer vision tasks.

h. Detectron2.
Figure 4. source:https://neurohive.io/en/news/fair-released-detectron2-new-object-detection-library/

Developed by Facebook AI, Detectron2 is a flexible and powerful framework for object detection. It is built on PyTorch and provides a wide range of pre-trained models and tools for researchers and developers to experiment with and deploy object detection models.

i. Hybrid Models.

Some recent object detection algorithms combine the strengths of convolutional neural networks (CNNs) and transformers, such as Vision Transformer (ViT). These hybrid models attempt to capture the spatial and contextual information in images effectively.

Part B: Regulations on Deep Learning Object Detection

1. Recent regulations for object detection algorithms
Figure 5. source: https://swisscognitive.ch/2022/05/20/european-union-eu-artificial-intelligence-act-ai-act-an-overview/

As of my last knowledge update in September 2021, several regions and countries were actively discussing and implementing regulations for AI systems, including those used in object detection. Please note that the regulatory landscape may have evolved since then.

2.European Union's AI Act.
Figure 6. Key features around EU AI Act.

 The European Union proposed the AI Act, a comprehensive regulation to govern the use of AI technologies. It addresses various aspects of AI, including object detection.

Figure 7. Key areas that regulations for AI systems in object detection address.

Several countries and regions were actively considering or implementing regulations for AI systems, including those used in object detection (September 2021).

Tags: