Purdue CNIT 581 ATR Fall 2018 – Object Detection Application for Physically Disabled People

This week, we kept polishing the detail of our SSD model for object detection. For test purposes, we did not use large dataset but using the pictures that we downloaded from google image. We utilized the google_image_download API which can help us download the image with key words. We can download maximum 100 pictures at a time. After downloading the pictures, we deleted the unwanted pictures such as wrong file types or wrong objects in the image. This week, we only download a few hundreds of images of cellphones and desks/tables, because we need eventually identify where the object will be. For instance, "a cellphone is on the table." The picture labelling process is very time consuming. We need to label each pictures and save an annotation xml file associated with the original pictures.

  We had our weekly meeting on Monday, and we finished our proposal report and searched the functional part for object detecting and Google speech to text API.   We have been testing a few solutions for object detection latterly. There are primary manners we have considered: 1. YOLO; 2. SSD. Before we dive into the implementation, we tried to find if we can use current datset that has labeled data for our target envirionment. We looked at the COCO dataset which is the most well-known datset for object detection.  In addition, we have been working on implement the YOLO and SSD model for object detection. We successfully implemented a YOLO pre-trained model on our laptop. We did some testing for our labled picture. As for SSD, we are still debating how to implement and use it in our project. In the end, we will select one method for the final project.

We finalized and submitted our preliminary proposal. Our new plan is to use web application to help the Physically Disabled People to detect object of their needs. The most updated schedule is listed bellow.