Our first client:
How did we end up collecting photos at trash dumps
The most common answer to the question of how to get your first customer is simple: your friends, your neighbors, or someone you can reach out easily. We must have done something wrong because our first client came from Chile. The first image dataset we collected was for a customer 17,000 kilometers away.
We received an email from a guy named Matías. First, it wasn't clear what his task was. He said he was working on a recycling project, and it looked like an image annotation order. During the Zoom call, it became clear that the client needed an image dataset. Afterward, we wrote the task description ourselves and sent it to the customer for approval.
Briefly, they required us to take images containing different categories of waste. Each piece had to be labeled using bounding boxes. Each image was supposed to have up to 5 objects classified into particular categories of waste (Glass, PET, Aluminum, Beverage Carton).
Briefly, they required us to take images containing different categories of waste. Each piece had to be labeled using bounding boxes. Each image was supposed to have up to 5 objects classified into particular categories of waste (Glass, PET, Aluminum, Beverage Carton).
The order didn't sound very complicated and wasn't very large. However, where can you get thousands of images that meet the customer's requirements without scraping the web? Very simple, go outside and take all these pictures.
The most surprising thing was the reaction of passers-by to our actions. People came up and asked why we were taking pictures of garbage. People buy soft drinks in hot weather, drink on-the-go and immediately throw empty cans and bottles right on the street. Was there going to be an article on dirty streets or a report to the mayor's office? Was the city administration going to strengthen the fight against waste on the streets? I couldn't tell them that some guys on the other side of the world needed it, so I told them it was needed to deal with our garbage problem.
It turned out that garbage photographer is such a respected profession that I was given access to restricted areas when I said that I needed to take images of garbage if they had it.
The most surprising thing was the reaction of passers-by to our actions. People came up and asked why we were taking pictures of garbage. People buy soft drinks in hot weather, drink on-the-go and immediately throw empty cans and bottles right on the street. Was there going to be an article on dirty streets or a report to the mayor's office? Was the city administration going to strengthen the fight against waste on the streets? I couldn't tell them that some guys on the other side of the world needed it, so I told them it was needed to deal with our garbage problem.
It turned out that garbage photographer is such a respected profession that I was given access to restricted areas when I said that I needed to take images of garbage if they had it.
Lifehack
If you want to infiltrate Area 51 and learn all about aliens, pretend you are taking pictures of trash!
We took images of trash in 9 cities from 4 countries, but some of our guys ran into bad luck. They lived in the center of clean cities. They had to drive to the outskirts and suburbs in search of garbage. Some even went to the city dump. Another way was to go to the beach at in the morning before the garbage was removed.
After a week of wandering through landfills, hidden, and not so hidden places, we had taken the required number of images. The next step was to get rid of low-quality photos. That sounds easy, but the quality of the images was not binary. Some of them were indeed of poor quality, but it wasn't easy to make a decision for some of the pictures. And then, we established a simple and obvious rule. If it is not clear whether the image is of sufficient quality, then it IS clear that the image is not of sufficient quality. We just deleted everything we didn't like and took more pictures. It was good that our smartphones still had a little charge left.
The last part of the preparation, the labeling, was pretty straightforward. Sometimes we have to collect a lot of metadata while shooting the datasets, but that was not the case this time.
Finally, we received the most welcoming feedback that you can get from the client: 'I checked the images, and they are perfect! Great work! Tell me about the payment, how should I pay you?'
Finally, we received the most welcoming feedback that you can get from the client: 'I checked the images, and they are perfect! Great work! Tell me about the payment, how should I pay you?'
After all that, we had 1,115 images containing 2,928 objects.
PET 875
Aluminum 732
Glass 881
Beverage carton 440
That dataset was very small compared to our subsequent works and usual demands for model training, but its main advantage was the strict requirements for that customer's needs. Whatever requirements the customer sets, we always satisfy them by creating a dataset from scratch.
PET 875
Aluminum 732
Glass 881
Beverage carton 440
That dataset was very small compared to our subsequent works and usual demands for model training, but its main advantage was the strict requirements for that customer's needs. Whatever requirements the customer sets, we always satisfy them by creating a dataset from scratch.