Electrical and Computer Engineering

Project PLANTS

Roots and Routes

Project PLANTS project image
Here is a picture of the team with our sponsor and instructor!

Project Description:

The overall project was to help UMD students be more enthusiastic about plants on campus by being able to identify specific plants by genus, species, etc. We started by creating an AI model which the team used a model by Inaturalist competitions 2nd place winners -metaformer. We later decided on a backup approach through location based information by incorporating AR and location of each plant on campus. Because we ran into problems with the natural offset of GPS inherit to phones we looked into other options like VPS but those all costed money. As the team continued to train we reached 72% accuracy  and by incorporating location as well the AI model became much more accurate. By allowing the model to use database location of closest plants the model became much more accurate and we even tested during maryland day.

 

The project set out to get UMD students genuinely excited about the rich variety of plants scattered across our campus by giving them a hands on way to learn each specimen’s genus, species, and common name. We began by building a machine learning model using MetaFormer, the runner-up in the iNaturalist competition and trained it on thousands of labeled plant images. That model alone eventually achieved about 72 percent accuracy in the lab.

To make our system more robust in the real world, we layered on a location-based fallback: we collected precise GPS coordinates for every tree, shrub, and flowerbed on campus, and tied those coordinates to simple web-AR markers. As soon as someone held up their phone, the app would pull in the closest handful of plants and display floating markers right where those plants should be. Of course, consumer-grade GPS on smartphones always introduces a 3–5 meter “drift,” especially when you’re standing under tree canopies or near buildings. We explored higher-precision services like Visual Positioning Systems (VPS), but the free tiers didn’t account for the offset.

 

Meanwhile, we kept iterating on our AI. By combining camera-based recognition with a list of nearby database candidates, our effective accuracy climbed well above 72 percent. During Maryland Day, we set up live demos around McKeldin Library and the Plant Sciences Building and received great feedback. They could point at a patch of flowers, tap the button, and see exactly the plant they were looking at. 

 

Advisor/Instructor:

Dr. Purtilo

Sponsor:

Dr. Neel

Team Members:

Ryan Ding Electrical and Computer Engineering
Andrew Le Electrical and Computer Engineering
Omar Malash Electrical and Computer Engineering
Nhat Nguyen Electrical and Computer Engineering
Scott Tran Electrical and Computer Engineering
Clark Wishard Electrical and Computer Engineering
Jiayi Wu Electrical and Computer Engineering

Table #:

F5
Back to Top