The deep learning framework KotlinDL 0.4 is available. The innovations and changes include a high-level API for what is known as “pose detection”, which is still described as experimental, with the help of which the posture (pose) of a person in an image or video can be recognized.
“Pose Detection” detects how a person moves
The new deep learning API for Kotlin is available for download from the Maven Central repository. The creators behind this library have introduced a whole range of new models and list the still developing “pose detection” as the most important innovation in this release of the framework.
This pose detection uses an ML model to detect a person’s posture in an image or video. To do this, the software determines the spatial positions of the most important body joints (key points).
The Kotlin Deep Learning API can detect and mark postures in images and videos.
JetBrains has developed the MoveNet family of detection modes with a new pose detection API in KotlinDL. According to the developers’ blog article, MoveNet is a fast and accurate model that recognizes 17 key points on the body. The model is offered on ONNXModelHub in two variants: MoveNetSinglePoseLighting and MoveNetSinglePoseThunder. MoveNetSinglePoseLighting is intended for latency-critical applications, while MoveNetSinglePoseThunder was developed for applications that require high accuracy.
New models to solve the object detection problem
Up until this version 0.4, the KotlinDL ModelHub contained only one model (SSD) that was suitable for solving the object detection problem. With the new version, the JetBrains developers have started to gradually expand the library’s capabilities to solve the object detection problem. To do this, they introduced a new family of object detectors called EfficientDet. They are said to be able to achieve much better efficiency than previous object detectors over a wide range of resource constraints.
Better callback support
In previous versions of KotlinDL, callback support was fairly basic, but not fully compatible with the Keras deep learning library. As a result, users had difficulty implementing their neural networks, building the custom validation process, and monitoring the neural network training.
A callback object was previously passed during compilation and was unique for each stage in the model life cycle. However, the model compilation can be in different places in the code than Fit/Predict/Evaluate. This means users need to create different callbacks for different purposes, which is now possible with KotlinDL 0.4. The JetBrains developers provide a detailed example for the support of multiple callbacks on GitHub.
The article in the Kotlin blog already offers a very comprehensive overview of what’s new in KotlinDL 0.4. Developers who need further and more in-depth information can find it in a readme file and on the KotlinDL project page. There you will also find a “Quick Start Guide” that provides detailed information on the basic and advanced functions of the library and goes into more detail on the topics mentioned in the blog post.
To home page
#Deep #Learning #Framework #KotlinDL #mit #PoseDetection #API