AIチップ

Use your own instance classification or use our ready-made ResNet-50-based classification application, trained on ImageNet

Depth estimation is prominently used for assisted driving, enabling smart automotive cameras to turn a 2D video input into a 3D to sense how close and far objects are. Use your own network or use the pre-built MonoDepth 2 model, pre-trained on KITTI depth.

Today, lane detection is a must-have in ADAS systems in vehicles of all classes and a basic feature in intelligent automotive cameras. Use your own network or use the pre-built PolyLaneNet-resnet34 model, pre-trained on Tusimple.

Commonly used for ADAS applications, in automotive cameras , semantic segmentation enables the vehicle to decide where the road, sidewalk, other vehicles and pedestrians are. The pre-built TAPPAS semantic segmentation app demonstrates Hailo-8's high compute capacity by processing a Full HD input video stream in real-time. The app is based on an FCN16-resnet18 network pre-trained on CitysScapes.

This useful app recognized moving people, determining the direction of their movement and tracking them. Uses for this function are endless, from human and vehicle traffic management for Smart City, to tracking customer traffic in Smart Retail operations. Use your own tracking network or use the pre-built FairMOT-regenetx-800mf model pre-trained on MOT-16 dataset.

Face detection apps are increasingly adopted for security and authentication tasks in public safety, commercial and consumer applications. From detection of a person, his gender, whether he or she is wearing a face mask or protective gear or not, and to recognition of the individual’s identity for such applications as access control. TAPPAS offers several apps, modes of operation and neural networks for detection and recognition. These include: ArcFace with a Mobilenet-V2 or ResNet-50 backbone (trained on the LFW dataset) 5-landmark detection using the RetinaFace architecture, trained on WIDER-Face.

Pose Estimation apps can be a significant building block for commercial and consumer applications. From recognizing emergency situations at home or on the factory floor, to analyzing customer behavior for better business outcomes. Use your own pose estimation network or use the pre-built openPose-regnetx-1.6GF model pre-trained on COCO.

Use your own instance classification or use our ready-made ResNet-50-based classification application, trained on ImageNet

Object detection is used in various commercial and consumer applications across industries. It is a fundamental functionality, used in complex applications such as traffic management for Smart City, product sorting in Industrial Automation and Smart Retail and home security, to mention a few. Use your own object detection network or leverage pre-built models, including:​

・YoloV5S/M
・CenterNet 18/50
・MobileNet v1 SSD with on-chip NMS

All networks are trained on the COCO dataset. Unique capabilities include: Tiling​Exploit Hailo-8's high throughput to process high-resolution images (FHD, 4K) by breaking the input image into smaller tiles. Processing high-resolution images is especially beneficial for scenes with many small objects, such as public safety applications at crowded locations, crowd analytics for Retail and Smart Cities and much more.

Instance segmentation identifies, classifies and outlines the shape of different objects. It is seminal in applications like robotics, where the machine needs to recognize the precise special position of the object to navigate safely and effectively. Use your own instance segmentation network or use the pre-built Yolact-regnetx-800mf model pre-trained on COCO.