Models and Pipelines

Cordatus uses models to detect and categorize objects in video streams. You can choose from Cordatus's ready-to-use computer vision models, or integrate your own custom models if the available options don't meet your needs. Currently, Cordatus supports Object Detection and Classification models. Semantic Segmentation and Pose Estimation models will be added soon.

NOTE: Classification models can only be used within pipelines by connecting to an object detection model; they cannot operate independently.

Cordatus also enables users to develop custom AI models by creating pipelines that easily combine the platform's ready-to-use models with users' custom models. This feature allows users to develop AI solutions specifically tailored to their needs.

Pipelines are one of Cordatus's core features, enabling multiple models to work sequentially. For example, in a vehicle detection model, users can first detect vehicles and then classify the vehicle's make and type using additional models. This demonstrates Cordatus's versatility and flexibility in addressing various use cases.

Models Page

---

Ready-to-Deploy Models & Pipelines

Cordatus currently offers two model architectures: Object Detection and Classification.

Ready-to-Deploy Object Detection Models

Model	Description
TRAFFICCAMNET	The TrafficCamNet model detects one or more physical objects in images captured by a stationary camera and categorizes them into four different categories. It provides bounding boxes and category labels for each object. Four categories : cars, people, road signs, and two-wheelers.
DASHCAMNET	The DashCamNet model identifies one or more physical objects in images captured by a moving camera. It provides bounding boxes and category labels for each object. Four categories : cars, people, road signs, and bicycles.
LPDNET	The LPDNet model identifies one or more license plate objects in an image. For each detected object, the model returns a bounding box and the lpd label.
PEOPLENET	The PeopleNet model identifies one or more physical objects categorized into three categories in an image. It includes bounding boxes and category labels for each detected object. Three categories : people, bags, and faces.
FACENET	The FaceNet model detects one or more faces in an image. It creates bounding boxes and assigns category labels for each detected face.
FACENET-IR	The FaceNet-IR model excels at detecting one or more faces in provided images or videos, specifically optimized for infrared (IR) images. Compared to the PeopleNet model, it delivers superior results in detecting large faces, such as those in webcam images.

Ready-to-Deploy Classification Models

Model	Description
GENDERNET	The Gendernet model classifies detected individuals by gender: male or female.
AGENET	The Agenet model excels at estimating the age of detected individuals between 0 and 100.
EMOTIONNET	The EmotionNet model categorizes human emotions detected by detection models into six different categories : Neutral, Happy, Surprise, Squint, Disgust, and Scream.
VEHICLETYPENET	The VehicleTypeNet model classifies car images into six different vehicle types : Coupe, Sedan, SUV, Van, Large Vehicle, and Truck.
VEHICLEMAKENET	The VehicleMakeNet model classifies car images into 20 popular car brands : Acura, Audi, BMW, Chevrolet, Chrysler, Dodge, Ford, GMC, Honda, Hyundai, Infiniti, Jeep, Kia, Lexus, Mazda, Mercedes, Nissan, Subaru, Toyota, and Volkswagen.
LPRNET_US	The LPRNet (License Plate Recognition Network for US) model predicts sequences of license plate characters detected by LPDNet, enhancing the accuracy and reliability of license plate recognition tasks.

Ready-to-Deploy Pipelines

Demographic-Analytics

The Demographic Analytics pipeline is designed to make comprehensive demographic predictions for detected individuals. Starting with FACENET, the pipeline accurately detects faces. The pipeline then seamlessly integrates GENDERNET for gender classification, EMOTIONNET for emotional state analysis, and AGENET for accurate age estimation. Together, these components create a powerful pipeline that provides detailed demographic analytics for comprehensive understanding of the target audience.

Demographic Analytics

Plate-Recognition

The Plate Recognition pipeline is designed for precise identification of vehicles and their license plates, providing advanced license plate recognition (LPR) capabilities. The pipeline begins with TRAFFICCAMNET for accurate vehicle detection, followed by LPDNET for license plate detection. Finally, the pipeline uses LPRNET_US for license plate recognition, ensuring a seamless and efficient process from vehicle identification to plate reading. This integrated approach provides reliable and accurate license plate recognition for various applications, from security to traffic management.

Plate Recognition

NOTE: Additional ready-to-use models and pipelines will be introduced soon.

---

Using the Interface

The Models page is a central location where models and pipelines are displayed together. Each item is labeled with its type:

• Model tag: For standalone models
• Pipeline tag: For pipelines combining multiple models

Interface Features

• Sorting: You can organize models and pipelines alphabetically using the sorting options at the top of the page.
• Search: Use the search box in the upper right corner of the screen to quickly find specific models or pipelines.
• Add New Model/Pipeline: Click the New Model button in the upper right corner to add a new model or pipeline.

---

Adding a Custom Model

If the available ready-to-use models don't meet your needs, you can add your own custom object detection and classification models to Cordatus. These models will be private and accessible only to you.

Supported Model Architectures

Cordatus supports the following training sources and model architectures:

NVIDIA TAO Toolkit Models

Models trained with the NVIDIA TAO Toolkit can be uploaded directly.

TensorFlow Object Detection API

Models trained with the TensorFlow Object Detection API are supported.

YOLO Model Family

Cordatus allows you to upload YOLO models you've trained yourself. To upload a YOLO model, you only need two files:

• Model File (.pt): YOLO model file in PyTorch format
• Label File (labels.txt): Label file containing one class name per line

💡 Tip: No additional configuration or conversion process is required for YOLO models. Cordatus automatically converts the .pt file to a TensorRT engine and optimizes it.

The following YOLO architectures are supported:

Model Family	Sizes	Description
YOLO11	N, S, M, L, X	Ultralytics' latest YOLO version. Improved accuracy and speed balance.
YOLOv10	N, S, M, L, X	Faster inference with NMS-free design.
YOLOv9	S, M	Improved learning with Programmable Gradient Information (PGI).
YOLOv8	N, S, M, L, X	Ultralytics' popular anchor-free architecture.
YOLOv7	Various	High performance with bag-of-freebies.
YOLOv6	N, S, M, L	Industrial model developed by Meituan.
YOLOv5	N, S, M, L, X	Widely used, mature and stable implementation.
YOLOv5u	N, S, M, L, X	Anchor-free version of YOLOv5.
YOLOX (Coming Soon)	S, M, L, X	Megvii's anchor-free YOLO variant.
YOLO-NAS (Coming Soon)	S, M, L	Deci AI's Neural Architecture Search model.
YOLOR (Coming Soon)	X	Unified network approach.

Other Supported Architectures

Model Family	Sizes	Description
Gold-YOLO (Coming Soon)	N, S, M, L	Improved feature fusion mechanism.
DAMO-YOLO (Coming Soon)	S, M, L	High-performance detector from Alibaba DAMO Academy.
D-FINE (Coming Soon)	N, S, M, L, X	DETR-based fine-grained detector.
RT-DETR (Coming Soon)	L, X	Baidu's Real-Time Detection Transformer model.

Model Size Guide

Size	Parameters	Recommended Use
N (Nano)	~3M	Edge devices, Jetson Nano, low power
S (Small)	~9M	Embedded systems, Jetson Xavier NX
M (Medium)	~25M	Balanced performance, general use
L (Large)	~43M	Applications requiring high accuracy
X (Extra Large)	~68M	Maximum accuracy, powerful GPUs

NOTE: YOLO models are typically pre-trained on the COCO dataset (80 classes). You can retrain the model through transfer learning for your own custom classes.

1. Adding a TAO / TensorFlow Model

1. Go to the Models page from the left menu of the Web Application.
2. Click the Add Model button. This opens the Add New Algorithm or Deep Learning Models modal.
3. Fill in the following information in the modal:

Field	Description
Model Label	Enter a name for your model.
Algorithm Type	Select the type of model you trained: object detection or classification.
Train Source	Select the application the model was trained with (TAO Toolkit or TensorFlow).
Model Key	Required only for models trained with TAO Toolkit; enter the model key used in the training configuration.
Model Format	Required only if TensorFlow Object Detection API is selected as train source. Select your model's file export format (ONNX, TF SavedModel, TF GraphDef, TF Checkpoint, or TF Keras).
Channel Order	Select your model's channel order as channel-first (NCHW) or channel-last (NHWC).
Channel Width and Channel Height	Enter your model's input channel dimensions.
Model Files	Upload your model's exported files (only zip format accepted).
Label File	Upload a file containing model labels on each line (only txt format accepted).

4. Click the Save and Upload button to complete the process.

NOTE: To ensure successful deployment, verify that your training configuration matches the configurations added to Cordatus. Any inconsistency in model configurations may lead to deployment errors. Carefully review and align these configurations to ensure a smooth integration process.

2. Adding a YOLO Model

Adding a YOLO model is quite simple. You only need two files:

Weights (.pt) and labels (.txt)

1. Go to the Models page from the left menu of the Web Application.
2. Click the Add Model button.
3. Fill in the following information:

Field	Value
Model Label	A name for your model (e.g., "Helmet Detector")
Algorithm Type	Object Detection
Train Source	YOLO
Model Size	N, S, M, L, or X (Nano, Small, Medium, Large, Extra Large)
Model Files	ZIP file containing your.pt file
Label File	labels.txt file (one class name per line)

4. Click the Save and Upload button.

Example labels.txt file:

person
car
truck
bicycle
motorcycle

NOTE: Cordatus automatically converts the YOLO .pt file to ONNX and then to TensorRT engine. This process occurs during the first deployment and may take some time.

---

Once the upload process is complete, Cordatus will process your model files and securely save them to your account. You can then start a job using this model. For more information, see the Jobs section.

---

Creating a Custom Pipeline

Creating pipelines with Cordatus is an effortless process that requires no coding. Users can easily design their pipelines by dragging and dropping desired models using the intuitive Cordatus Model Composer.

Steps to Create a New Pipeline

1. Go to the Models page from the left menu.
2. Click the New Model button in the upper right corner.
3. Select the Pipeline option from the dropdown menu.
4. The Model Composer will open. Start your pipeline with the Dummy Stream. This represents your camera stream and serves as the starting point for all pipelines.
5. Find the models you want to include in the left sidebar and drag them to the Model Composer. The Dummy Stream and first model will automatically connect.
6. To create connections between models, drag a model's output and draw a line to connect it to another model's input.
7. To remove a connection between models, drag the input or output part of the connection line and drop it in the black area.
8. To remove a model from the pipeline, right-click on the model's header and select the Delete option.
9. Click the Save Pipeline button, enter a name for your pipeline in the Pipeline Label field, and click the Save button.

Once your pipeline is created, it's ready to deploy. For details on starting a job using this pipeline, see the Jobs section.

NOTE: For optimal performance and efficiency, it's recommended that your pipeline start with an object detection model and continue with a classification model.

---

Which model size should I choose when uploading a YOLO model?

Size	Recommended Use
N (Nano)	Edge devices, Jetson Nano, low power consumption
S (Small)	Embedded systems, Jetson Xavier NX
M (Medium)	Balanced performance, general use
L (Large)	Applications requiring high accuracy
X (Extra Large)	Maximum accuracy, powerful GPUs

Can a classification model be used standalone?

No, classification models must be used within a pipeline connected to an object detection model.

What's the difference between a Model and a Pipeline?

• Model: A standalone AI model (e.g., detecting people with PEOPLENET)
• Pipeline: A flow where multiple models run sequentially (e.g., Face detection → Gender prediction → Age prediction)

How many models can I use in a pipeline?

There's no technical limit, but for performance it's recommended to start with an object detection model and continue with classification models.

---

Copyright © 2025 Cordatus.