CircularNet

Instance segmentation models for identification of recyclables on conveyor belts.

We provide retraining and fine-tuning utilities, but if you're interested in partnering more closely with us reach out to [email protected]

Overview

Circularnet is built using RF-DETR, a vision transformer model that includes both object detection and instance segmentation, which is a deep learning model for instance image segmentation, where the goal is to assign instance level labels (e.g. person1, person2, cat) to every pixel in an input image.

Model Categories

• Material Type: Identifies the material type (metal, paper etc) of an object. For plastic, resin types are also identified (HDPE, PET, LDPE, etc).
• Material Form: Categorizes objects based on the form factor (cup, bottle, bag etc)
• Example inference label: Plastics-PET_Bottle

Latest model

Single unified model that performs material type and form detections

Model categories	Model backbone	Model type	GCP bucket path
Material Type & Form	Vision transformer	onnx model	click here

Full Documentation

The full documentation, covering everything from how to choose and install a camera to how to prepare and make use of the model is here. Below, we also provide a quicker guide for running inference using a GCP VM, assuming you already have a working camera taking pictures.

End to End Cloud Deployment Guide

End to end deployment involves three key steps:

1. GCP GPU VM creation

2. Code configuration

3. Results analysis

We will go through each one of them in details below

[A] Prerequisite - Create VM instance:

Create a Google cloud account and a T4 GPU enabled VM:

• Create VM in GCP Cloud

[B] Code Setup - Clone and start the pipeline

Run the following commands mentioned in each step on the SSH-in-browser window of your VM instance in Google Cloud

Step 1:

• Clone the repository

Step 2:

• Start the server

Step 3:

• Run the prediction Pipeline

For more details: Click Here

[C] Setup Dashboard - Visualize results

For reporting purposes and to analyze image categories, we need to set up and connect looker dashboard with BigQuery table:

• Prepare and analyze images

Authors and Maintainers

Umair Sabir - Primary developer Vinit Ganorkar - Primary developer Ethan Steele - Collaborator Sujit Sanjeev - Product Manager