https://github.com/abhi1thakur/tez
Goal
To train single object detection models from scratch on commmonly available datasets and make them available them in small enough packages so that they can be deployed easily. Add deployemnt scripts with FastAPI, TfJS, TFLite and TFMicro. Also provide option to train object detetcors using any subset of categories available in these datasets
Available detectors
- Person
- Use cases include people counter pedestrian detection
- Automobile
- For lience plate detection
- Animals
- Objects
Datasets
PASCAL MS COCO GOOGLE OPEN IMAGES
Models
YOLO YOLO v2 YOLO v3
mobile net SSD retina Net
FastAPI Example
Tf Micro Example
Results Comparison table with pretrained models
Resources
Blogs
- Lesson 9 SSD FASt AI
- FAST AI Part 2 lesson 9 wiki
- Concepts in object detection
- Deep Learning for Object Detection: A Comprehensive Review
- Understanding SSD MultiBox — Real-Time Object Detection In Deep Learning
- The Modern History of Object Recognition — Infographic
- Understanding anchors
- A guide to receptive field arithmetic for Convolutional Neural Networks
- The effective receptive field on CNNs
- Receptive Field Calculator