Applying Deep Learning for Object Detection from a Telecom Perspective
summary
VodafoneZiggo builds an image pipeline that collects 47,000 images, prunes ~10k with OpenCV variance of Laplace and HOG cosine checks, labels ~8,000 then augments to ~32k, and trains a YOLOv3 detector initialized from Darknet53 weights. Toolchain includes MS VOTT, AWS Ground Truth then CVAT, TensorFlow and planned TensorFlow Lite mobile deployment. End-to-end takes about nine months, model reaches precision 87% and runs as an AWS service with plans for YOLOv3 → TensorFlow Lite deployment.
AI-written summary · about
chapters (18)
- 0:00Deep learning in telecom
- 1:10Object detection at Vodafone Ziggo
- 3:16Telecom infrastructure and Giganet
- 5:06Self-installation challenges and churn
- 7:36AI-powered app checks solution
- 8:40Data collection for training images
- 10:05Technician app for image capture
- 11:30Image quality control and cleaning
- 12:53Labeling tools and workflow
- 15:43Data augmentation for training
- 17:29Model training and performance
- 18:33Deploying model to production
- 21:01Production-ready ML engineering
- 22:44Modular ML pipeline architecture
- 24:10Storm for training and inference
- 25:36Customer impact of the solution
- 25:59Live AI installation check demo
- 27:45Q&A on labeling and transfer learning
related talks
38mDeep Learning for detection on a phone
01h 15mTrain Object Detection with Small Datasets
46mML Use Cases at Think Big Analytics
21mUsing Computer Vision to Solve Interesting Problems
01h 14mFoundations of Object Detection from an Expert's Overview
34mCard detection using YOLO on Android
13mMLOps Pipelines to Deliver Tf.Lite Models to Distributed Compute Pools at the Edge
12mCascade models in computer vision to boost accuracy and performance