Case Study: DeepPlastic and YOLOv5

Plastic is suffocating marine wildlife: every minute, two truckloads of plastic are dumped into our oceans, equating to over 10 million tons a year. DeepPlastic scientists state that this marine plastic poses societal threats to the "marine environment, food safety, human health, eco-tourism, and contributions to climate change."

To combat this, this team of researchers and engineers have been investigating how computer vision can eliminate plastic in our oceans.

With deep learning technology, DeepPlastic researchers have developed an approach that uses autonomous underwater vehicles (AUVs) to scan, identify, and quantify plastic located just below the surface of the ocean where light can still penetrate through, or the Epipelagic layer.

“Our goal was to have a very tiny model with a very fast inference speed that can be used to detect plastic.”

Jay Lowe, Machine Learning Researcher

The DeepPlastic team trained two small and precise models, YOLOv4 and YOLOv5, allowing for real-time object detection. These models were trained on the DeepTrash dataset, which consisted of:

• 1900 training images, 637 test images, 637 validation images (60, 20, 20 split)
• Field images taken from Lake Tahoe, San Francisco Bay and Bodega Bay in CA.
• Internet images (<20%) taken by scraping Google Images.
• Deep Sea images from JAMSTEK JEDI dataset

How Autonomous Underwater Vehicles (AUVs) Work

An AUV is a robot that travels underwater. They are slow vehicles that can glide freely to the ocean depths and back to the surface. A deep learning model must be installed into the AUVs for them to be able to identify and collect plastic underwater. AUVs can be deployed in three easy steps to detect plastic underwater.

1. Install a Deep Learning Model Into an AUV

2. Scan the Ocean

3. Identify Plastic

The Problems With Previous Deep Learning Models & Ocean Cleaning

The DeepPlastic team tested out several deep learning models such as YOLOv4 and Faster R-CNN on AUVs. However, researchers were presented with a range of challenges that made ocean cleaning problematic.

Limited Accessibility to Researchers Stalled the Team

Without any deep learning experts on the team, researchers were prevented from getting the most out of the deep learning models.

Slow Inferencing Speed Weakened Plastic Detection

Inferencing is how fast the AUV can recognize plastic. With YOLOv4 and Faster R-CNN, AUVs were not as effective in detecting plastic, blighting their ability to clean the water.

Low Accuracy in Object Identification

YOLOv4 and Faster R-CNN only had an average of 77%-80% success rate when identifying plastic.

Poor Detection Confused Corals With Plastic Objects

When using Faster R-CNN, 3-5% of corals were identified as plastic by AUVs, which was below the acceptable standard.

Greater Power and Accuracy With YOLOv5

When switching to YOLOv5, the researchers saw an immediate transformation. Accuracy was augmented, speed was maximized, and the simplicity of YOLOv5 made it accessible to everyone on the team.

20% Faster Inference Speed On Average When Compared To Faster R-CNN

93% Precision Rate

Less Than One Hour To Set Up YOLOv5

Increased Accessibility Allowed Researchers to Get the Most Out of YOLOv5

There were several aspects of YOLOv5 that allowed the team to easily work with it, based on the simple step-by-step process we’ve established on the repository.

• Downloading the repository was straightforward
• All documentation was organized in a clear, easy-to-follow manner
• Simplified model training
• Manual result-checking

Higher Inferencing Speed Maximized Ocean Cleaning Efficiency

YOLOv5 presented 20% faster inference speeds than Faster RCNN, processing an average of 1 image in 9 milliseconds. As a result, AUVs were able to detect floating plastic at a faster speed, which increased the amount of plastic captured and overall project efficiency.

Enhanced Accuracy in Precision Rates

Precision rates were at an 85% average that sometimes went as high as 93%. This is a jump from the 77-80% average seen with previous models.

Greater Ease of Use Allowed Researchers

Setting up YOLOv5 was both a seamless and effortless experience for the researchers. Users were guided A-Z throughout the entire set-up process allowing the team to get started with YOLOv5 in less than an hour.

Improved Versatility Allowed the Researchers to Apply YOLOv5 to Different Water Environments

Within a couple of days, using a small data set of 3000 images without augmentation, the group was able to train the AUVs to work in lakes and rivers. Despite murky water and other poor conditions, the AUVs trained on YOLOv5 could still detect and identify plastic with high accuracy.

“We were looking for an object detection algorithm that produces high accuracy and is extremely fast. The oceanic environments that we work in are harsh, rough terrains. YOLOv5 delivered on all fronts as the best object detection model we could’ve used.

“We love using YOLOv5 since it’s so easy to set up and use, and it has been producing the results that we’ve wanted consistently.

“For any future models that we will deploy, we will be looking at YOLOv5 as our first choice without a shadow of a doubt.”

Gautam Tata, Machine Learning Researcher

‍

Check out the DeepPlastic repository, published paper, and video recap.

‍