Production Object Detection - A Journey of Training, Building and Deploying CV models
Computer Vision as a field has changed manifold in the past few years. Researchers publish their papers and at times their code for the latest algorithms, but the challenge for the industry remains in applying that research to their processes.
Customising a company’s proprietary data for the research models, implementing their code, and training models is the first big hurdle. Then comes the part where we have to test and release these latest models to production.
In this talk we will go through a project where we did exactly the above at Here Technologies. The audience will learn abot the main issues we faced, how we overcame it and other best practises, including optimising AWS infrastructure for Machine Learning DevOps.
- Overview of object detection approaches
- Training Data Prep - Including handling data on the cloud
- Data collection - Sampling
- Annotation and review approaches - human, automated
- Actually training the model - hardware/cloud/best practices
- Troubles with large data sets, how to deal with issues when you hit the limit of state of the art hardware
- Evaluation of the model results
- Double checking the evaluation - blind test dataset
- Release and integration with systems
- Deployment and Infrastructure
None as such. For the talk, the only pre req is basic knowledge of machine learning terminology.
I’m an engineer involved in computer vision and robotics since 5+ years. I have worked on various computer vision and data science projects including an autonomous soccer playing humanoid(acyut.com), OCR(text extraction/transcription) and object detection models. As a computer vision and data science practitioner who has faced and overcome challenges in production systems, it would be great to share some of that knowledge for the benefit of the community.