Shelf Object Recognition Model

SSD generated model capable of identifying 5 distinct micro categories of sodas

This project presents an approach for identifying and locating products on supermarket shelves with a focus on distinguishing different flavors of the same product, whose problem remains a challenge due to the dynamism of the context in which the product is inserted and the frequent change in packaging. Whose tests resulted in an accuracy of 83.6% and precision of 87.0%.

Link to the paper

This app main purpose is testing if the generated model is capable of identifying this 5 classes, ["coca-original-350-sem-acucar", "coca-original-600", "coca-original-310", "campo-largo", "coca-original-350"] in a video recorded in a context containing shelf products (e.g., supermarkets, convenienve stores, etc.).

Prerequisite

---

• Node >= 8.x
• NPM >= 6.x

Project structure

---

• public/model_web: Contains the trained model

• public/video: Contains the video that will be used for testing the model.

Run the project

---

First install the packages needed:

$ npm i

Command to start the application :

$ npm start

Then and access it under http://localhost:3000

Test it with your own video

---

• First install ffmpeg package on your machine (Quick tutorial)

• Run the convert.sh script passing the video you want to convert to HLS (m3u8) type

  $ ./convert.sh <VIDEO_PATH>

• Place the generated files in the folder public/video

---

Credits to @bourdakos1 who created the base script for the web video stream app.