Half of the week was spent on the Design Review. There were issues related to my part specifically that were addressed in our feedbacks. First of all, there was a concern related to how we would customize the Yolov7 model to account for human objects and tune out non-human objects. Basically, the current Yolov7 model will output the coordinates of bounding boxes for every object in the frame and also map each set of coordinates to a specific object. The first column in output text files specify what object it is so all I’ll have to do is to filter out the output data by selecting only rows of data that have “person” in the their first columns. Another concern is related to our wait time algorithm. The current algorithm cannot guarantee a margin of error of 1-2 minutes because if the wait time for 1 person is within 10-30 seconds, then the margin of error will be way bigger than 1-2 minutes when we multiply the wait time by the number of people in the line. The variation of wait time for 1 person depends on what they order (e.g, latte, hot chocolate, sandwich, etc.). Since our goal for the MVP is to have our system work perfectly well in one dining location, we can collect data about the wait time for each order over a certain time frame and then perform a linear regression on the collected data to find the average wait time for each kind of order. However, to estimate the wait time of the line, this will require each person in the line to provide information on their order on the web application, which is impossible because we don’t want to make users unwillingly use the application. Therefore, we’ll need the cashier to provide the order information on our application. To collect data about each kind of order, we can take turns and collect the data at La Prima over a couple of days. To test our margin of error, we can have another feature specifically for testing reasons that let the system know when we make and receive the order. Then, we can easily compare the calculated wait time and the actual wait time and see how different our estimated result is.

Again, you can see my updated code here. I will update more on the design, test results on there after our meeting tomorrow (02/26/2023). You’ll see a lot of progress on there by then as I’m currently fixing some bugs. My progress is a little bit behind because there are adjustments to make to the design but I should have enough time to catch up before spring break because I have more free time slots next week. I hope to finish the customized code for object detection by next week and have the code for linear regression model and wait time calculation mostly done.

For this week, I’ve been spending time on the design aspect of my algorithm before actually implementing them next week. The goal for this following week will be to implement the YOLO algorithm for object detection and have it work on images with 1 to multiple objects. I’ve downloaded YOLOv7 and I’ve been looking at a lot of different resources to get familiar with it before writing actual code.

Here are some resources that I’ve found about YOLOv7 that will be useful for the project (and particularly the future design presentation):
1. https://github.com/WongKinYiu/yolov7

2. https://learnopencv.com/yolov7-object-detection-paper-explanation-and-inference/

3.https://viso.ai/deep-learning/yolov7-guide/

4. https://blog.paperspace.com/yolov7/

5. https://www.kaggle.com/code/taranmarley/yolo-v7-object-detection

6. https://blog.roboflow.com/yolov7-breakdown/ 

These resources also explain the performance tradeoff between YOLO and other algorithms (especially the updated performance compared to its previous versions) and solidify my choice of using this algorithm more for object detection.

My progress is currently on schedule. Our group is meeting tomorrow (Sunday 02/12/23) and I’m expecting to put in a few more hours of work on finalizing the design and getting started on writing the code. By the end of next week, I am expecting to finish the code that can detect objects in an image.