It’s monsoon season in Korea. This means every day is either pouring down or incredibly gloomy. We aren’t letting that hold us back though! Good food, shopping, and visiting museums are all fun indoor activities. Plus, a little rain never hurt anyone.

Research Time

This week, the amount of tasks I was assigned to complete in the research lab increased a lot. I am working under my mentor Taeho Lee and his TA Viduu. They are both very knowledgeable in machine learning and are patient in teaching me how to use YOLO and YOLACT. This was the first week I was assigned tasks to complete on my own as Viduu went to Jeju Island to attend a conference. 

Before I continue, I will explain my research project more. I am assisting Viduu on one of three of his own research projects, this one is based on a similar one completed at Stanford. In this project, we are training the YOLACT algorithm to be able to identify and segmentate different objects. To do so, we need to create a large dataset for each object. Viduu has already completed the dataset and training for 45 objects, I am assisting with an additional 15 objects. These objects are common items you would find in an American convenience store — snacks, drinks, mouthwash, ect. 

Figure 1: Displays the station where I record the videos for the object image databases. On the top of the shelf, you can see the 15 objects I am working with. On the top shelf, you can see the recording station with adjustable lighting, 3 different cameras, and a turntable to rotate the object. On the bottom shelves, you can see all the objects Viduu has already completed the databases of.

After Viduu showed me how to create a dataset for objects, I was left to my own devices. The process goes as follows:

  1. Record a video of the object from various angles
  2. Run the video through a program that splits the video into a series of images
  3. Run the images through an algorithm attempts to remove the background from the object.
  4. Go through the output of the final algorithm and remove all images that have been poorly edited.

Unfortunately, I did encounter some problems while completing my assigned tasks. One is that the original recordings were all completed with a white background. This meant that if the object was white or pale, the background removal algorithm would often think that the object is part of the background and would remove large portions of the object. As a solution to this, we had to modify the recording station by covering it with darker paper and re-record the videos. 

Figure 2: Top left displays the original recording of the image, top right displays the output of the background remover algorithm. As you can see, a large portion of the object is missing from the top of the object. The bottom left shows the re-recorded version of the image, and the bottom right shows the near perfect output of the background removal algorithm.

An additional problem encountered is when an object has transparency, like a coke bottle. The background removal algorithm tends to not be able to detect that the object is continuous, and thus will cut the transparent section out and assume it is part of the background. We have not yet found a solution to this problem.

Figure 3: Left shows the photo taken of the object, right shows the output of the background removal algorithm. As you can see, the top of the object ends up separate from the rest of the bottle.

So far, I have gone through over 36,000 different images to complete the datasets, with many more to come! Next week, once we finally complete all the datasets, we will start training the object detection algorithm.

Excursions

  1. A Night Out

After the incredible hike last week, along with my increased workload, this week I was dead tired. Thus, beyond going to restaurants to eat dinner, I didn’t go out much during the week. On Friday evening, Nadia, Busa, and I all went to meet a friend at a hip-hop bar in Hongdae. On the way, we decided to stop for some street food. All street food here is delicious! 

Figure 4: Street food that we ate. They have corn dogs with cheese in them, various fried foods, topokki (rice cakes), and fish cakes.

 

There were quite a few foreigners in the area, which was surprising to see. Due to the pandemic, everything closes at 10 pm, so we didn’t stay out too late; we still had a lot of fun! 

  1. Day in the Museum 

On Sunday, the rainiest day so far, we decided to head over to the National Museum of Korea. It was incredible, they had a lot of artifacts and art; Korea has such a rich history. The museum is free, incredibly large, and very clean. If you visit here, make sure to leave a donation! The gift shop is also something great to check out, even if most of the things in there were too pricey for my pockets. They had intricate traditional golden crowns, expensive looking pottery, and large prints of art. Definitely worth checking out!

Figure 5: Ryoma(left), me, Nadia, and Busa(right) in front of the 10 story tall stone pagoda in the National Museum of Korea

Dr. Moser’s Workshop:

This week in Dr. Moser’s workshop, we worked on telling stories. Each of us told a story, meant to last 2 minutes. After we told the story, Dr. Moser and the other students all gave advice on what went well and how the story telling could be improved. After that, we practiced improvisational speeches. In this assignment, we were given a prompt and on the spot had to give a 2 minute response to it. My feedback for my talk was that I should provide more evidence and explanations to my claims and that I should be more confident with my responses. This advice was helpful and I feel as though I will be more prepared if the need ever arises. Next week we are giving Elevator Pitches, which are two minute introductions to yourself, your skills, and your goals. We will also be answering faux interview questions. 

Current Takeaways:

  • If there is something you need, don’t assume you can just buy it once you get to Korea! A lot of stores here are more specialized, meaning that you might need to look for a specific store to find what you need.