Optimize object detection model with data provided by DocketRun

Optimize object detection model with data provided by DocketRun

Optimize object detection model with data provided by DocketRun Technologies

DocketRun Technologies used ENAP platform to optimize the object detection models for edge devices. An end-to-end workflow solution was performed consisting of training, optimising and deploying model.

DocketRun:

DocketRun Technologies provide AI-powered solutions to companies across various sectors. Headquartered in Deshpande Startups (India’s Largest Incubation Centre), Hubballi, Karnataka and backed by the mentors like Santosh Huralikoppi and CM Patil, DocketRun has customers across segments like E-commerce, Retail, Manufacturing Plants, Educational Institutes, and so on, with customers like Maruti Suzuki, Tata Hitachi, Fabwoods, KLE University, Deshpande Educational Trusts, Alie Global, and others.

Challenge:

Increase the number of cameras supported by edge hardware while keeping the cost same.

ENAP:

ENAP is the Edge Neural’s AI platform. It is a unified, cloud-neutral and hardware-agnostic platform. It integrates workflow to automatically build optimised AI models that are smaller and efficient. The platform provides enterprises, AI engineers and start-ups a tool to deploy DNN (Deep Neural Networks) on resource constraint, low- power and cost-effective hardware as well as manage the models without compromising on the performance.

Models for object detection:

  • Yolov5
    Yolov3-tiny

Production Hardware:

  • NVIDIA Jetson Nano
  • NVIDIA Jetson TX2

Model Training Details:

Part 1

Part 2

Model

yolov5

yolov3-tiny

Epochs

300

300

mAP

74%

83%

Iterations

100000

Why ENAP?

DocketRun Technologies was using AI for one of the retail use case. They wanted to increase the number of cameras supported by edge hardware for analytics. With ENAP entire workflow solution was implemented starting from training the model, optimising it and then deployment on the edge device.

ENAP Training Docker:

The training for the provided dataset was executed using ENAP platforms training dockers. For this use case, EDGENeural’s team chose yolov5, yolov3-tiny models for object detection. The training dataset

Leave a Reply

Your email address will not be published.Required fields are marked *