What are the most recent developments in state-of-the-art machine learning (ML) systems, and deep learning models, for computer vision detection and classification? Specifically the architectures.

Part
01
of one
Part
01

What are the most recent developments in state-of-the-art machine learning (ML) systems, and deep learning models, for computer vision detection and classification? Specifically the architectures.

Hello, and thanks for your question asking what are the most recent developments in state-of-the-art machine learning (ML) systems, and deep learning models, for computer vision detection and classification. I have organized 36 links to research into categories. Below you will find a deep dive into my research, along with all the details as to how I came to this conclusion.

OVERVIEW
In order to answer your question I reviewed the information that you were already familiar with, and have then provided links to the research mentioned in each of the articles that you mentioned, and more. I have arranged these into categories, and have summarized which category appears most in this recent research.

FINDINGS
TENSORFLOW
This is the paper for TensorFlow described in the Google Research Blog article "Supercharge your Computer Vision models with the TensorFlow Object Detection API."

- TensorFlow Estimators: Managing Simplicity vs. Flexibility in High-Level Machine Learning Frameworks Aug 2017
This TensorFlow paper provides comparisons of various aspects of machine learning.

SSD
This link was included in the Google Research Blog article "Supercharge your Computer Vision models with the TensorFlow Object Detection API."

INCEPTION
This is an earlier paper about Inception which is described in the Medium article "An Intuitive Guide to Deep Network Architectures."

This is the more recent paper about Inception which is described in the Medium article "An Intuitive Guide to Deep Network Architectures."

RESNET
This is the ResNet paper described in the Medium article "An Intuitive Guide to Deep Network Architectures."

FASTER R-CNN
This is the paper for Faster R-CNN which is described in the Google Research Blog article "Supercharge your Computer Vision models with the TensorFlow Object Detection API."

This link was included in the Google Research Blog article "Supercharge your Computer Vision models with the TensorFlow Object Detection API"; the abstract says it is meant "to serve as a guide for selecting a detection architecture that achieves the right speed/memory/accuracy balance."

FULLY CONVOLUTIONAL NETWORKS
This is the Xception paper described in the Medium article "An Intuitive Guide to Deep Network Architectures."

This references the other Xception paper but draws separate conclusions.

This is the paper for MobileNets which is referenced in the Google Research Blog article "Supercharge your Computer Vision models with the TensorFlow Object Detection API."

This paper is about Max-Margin Object Detection and includes the Histogram of Oriented Gradients (HOG) and sliding window framework described in the Medium article "You Only Look Twice — Multi-Scale Object Detection in Satellite Imagery With Convolutional Neural Networks (Part I)" but concludes better results were obtained with the MMOD than without.

This paper deals with the convolutional neural network and geometric matching.

This paper deals with Irregular Convolutional Neural Networks.

This paper deals with deformable convolutional networks

This paper deals with convolutional neural networks, particularly Inception, as well as GoogLeNet.

This paper deals with improvements to object detection, particularly involving small objects which I believe relates to You Only Look Twice.

This is the paper for AlexNet.

This paper does not have a lengthy abstract but is by the same primary author as that of AlexNet and deals with improved training of CNNs.

This paper includes information on VGG-16.

Paper on fully convolutional networks.

CRFS
This is the most recent paper I found which discussed conditional random fields (CRF), included in the client's "Architectures" summary.

- Amortized Inference and Learning in Latent Conditional Random Fields for Weakly-Supervised Semantic Image Segmentation May 2017
This paper deals with CRFs and semantic image segmentation.

- 2D-3D Pose Consistency-based Conditional Random Fields for 3D Human Pose Estimation Apr 2017
This is an additional paper concerning CRFs.

Another paper concerning CRFs.

A further paper concerning CRFs.

GAN
This paper gives an overview of Generative Adversarial Networks.

This paper is about General Adversarial Models.

MISCELLANEOUS
This paper is about the COCO dataset, used in the detection models.

This paper is on the topic of a YOLO (You Only Look Once) model.

This paper discusses "pix-to-pix".

- GitHub
This is the GitHub page for the TensorFlow model zoo

This is a paper about the annual ImageNet competition and the advancements in object recognition which have resulted from it.

This paper includes information on EDeconvNet which is included in the client's "Architectures" summary.

- Comparing the Performance of L*A*B* and HSV Color Spaces with Respect to Color Image Segmentation Feb 2015
This paper deals with image processing, particularly HSV

This paper deals with "non-max suppression".

CONCLUSION
Overall I have found that the most recent developments in state-of-the-art machine learning (ML) systems, and deep learning models, for computer vision detection and classification are based on fully convolutional models. I have been able to provide links to 36 examples of recent research, and have arranged them into categories.

Thanks for using Wonder, please let us know if there's anything else that we can help you with!

Did this report spark your curiosity?

Sources
Sources