Model Selection and Options

We understand that your projects may have different accuracy requirements, and you may (or may not) be willing to trade accuracy for computational complexity. Therefore we have provided 28 model architectures (and more in the future) ranging from lightweight to highly complex models. Selecting the appropriate model for your use case can help you improve time efficiency and computational cost while balancing high accuracy. The impact is further detailed in Improving Model Performance.

Each model also has options for training parameters.

Advanced Options

There are also advanced options available for tuning specific hyperparameters.

For Video Classification Model Architectures

General Model Selection Tips

In general, models with larger dimensions on the end of the name imply more robust, complex models that are capable of taking in more data and thus more capable of learning more complex features for prediction. If you need higher accuracy and more complexity for your use case, then you should opt for higher dimensionality. However, if compactness and quicker training and inference is more important to you, then you should consider smaller dimensions.

When model names include names like ResNet, MobileNet, or InceptionV2, these represent different backbone models that are responsible for extracting image features such that the rest of the model can utilise these features for their own different processes. As a general trend of the same idea as the above paragraph, MobileNet is the most compact, ResNet is in the middle, and the number next to it, like 50 in ResNet50 indicates how many layers the model has, so the higher the number, the more complex it is. The most complex and robust is InceptionV2.

Model outputs differ based on the task that they are designed to solve. Datature currently offers models for the following tasks:

---

Models

Classification

YOLOv8-CLS

YOLOv8 is an extension and improvement upon previous versions of the YOLO family of algorithms that is known for their real-time object detection capabilities. YOLOv8 builds upon the concepts of the original YOLO algorithm, aiming to improve both accuracy and speed. It incorporates advancements such as feature pyramid networks, spatial attention modules, and other architectural improvements to enhance the detection performance. YOLOv8-CLS contains a classification head used for image classification tasks.

MoViNet

MoViNet is a family of CNN model architectures with a focus on efficient video recognition, particularly suited for mobile devices. Its design prioritizes computational efficiency while maintaining high accuracy, making it ideal for tasks like real-time video analysis on smartphones and tablets.

Object Detection

YOLOv9 [New!]

YOLOv9 is an extension and improvement upon previous versions of the YOLO family of algorithms that is known for their real-time object detection capabilities. YOLOv9 builds upon the concepts of the original YOLO algorithm, aiming to improve both accuracy and speed. It incorporates advancements such as Programmable Gradient Information and Generalized Efficient Layer Aggregation Network to enhance the detection performance.

YOLOv8

YOLOv8 is an extension and improvement upon previous versions of the YOLO family of algorithms that is known for their real-time object detection capabilities. YOLOv8 builds upon the concepts of the original YOLO algorithm, aiming to improve both accuracy and speed. It incorporates advancements such as feature pyramid networks, spatial attention modules, and other architectural improvements to enhance the detection performance.

RetinaNet

RetinaNet is a one-stage object detection model that utilises a focal loss function to address class imbalance in the training dataset. It has strong performances with dense and small scale objects.

FasterRCNN

Faster R-CNN introduces a Region Proposal Network (RPN) that shares convolutional features with the detection network, enabling low-cost region proposals. Further, they merge this RPN with Fast R-CNN (another single end-to-end unified object detection network for quick object detection) to achieve high quality, rapid object detection results.

EfficientDet

EfficientDet is another object detection model which uses optimizations and scalable tweaks rather than additional modules to improve object detection. This is a model that is advantageous due to its model efficiency and ability to scale adaptively.

YOLOv4 [DEPRECATED]

YOLOv4 is one of the newer one-stage object detection models running on DarkNet, which has achieved more improvements in the tradeoff in speed and accuracy of detection.

YOLOX [DEPRECATED]

YOLOX is an anchor-free version of YOLO, with a simpler design but better performance that makes several modifications to YOLOv3.

Semantic Segmentation

DeepLabV3 Semantic Segmentation

DeepLabv3 is a semantic segmentation architecture with improvements to handle the problem of segmenting objects at multiple scales.

UNet Semantic Segmentation

U-Net is a semantic segmentation architecture. It consists of a contracting path and an expansive path that consider both typical features from a convolutional network and a progressively upsampled feature map to improve detail.

FCN Semantic Segmentation

Fully Convolutional Network is a semantic segmentation architecture. It exclusively uses locally connected layers, such as convolution, pooling, and upsampling, and avoids the use of dense layers. This makes it faster to train and reduces parameter size.

Instance Segmentation

YOLOv8-SEG

YOLOv8 is an extension and improvement upon previous versions of the YOLO family of algorithms that is known for their real-time object detection capabilities. YOLOv8 builds upon the concepts of the original YOLO algorithm, aiming to improve both accuracy and speed. It incorporates advancements such as feature pyramid networks, spatial attention modules, and other architectural improvements to enhance the detection performance.

MaskRCNN

MaskRCNN is Datature's instance segmentation model designed for predicting segmentation masks using RCNN as the base model.

Keypoint Detection

YOLOv8-Pose

---