%20(1).png)
Background :
To kick off our series of Lixo Journal Club articles, we've chosen to present the paper ZeroWaste: Toward Automated Waste Recycling by Bashkirova et al. which was presented at the "2022 Conference on Computer Vision and Pattern recognition"!
This is a special type of article, since it does not improve a new architecture or loss calculation, but publishes a new dataset for object detection and segmentation on images of waste management facilities.
Why did we choose this article?
The ZeroWaste dataset is the first open-access dataset to provide images of sorting plants, in particular a conveyor belt containing paper ("fibrous" in industry parlance). These images differ from well-known datasets such as COCO or PASCAL VOC.
Indeed, the objects in the image are highly distorted, jagged, superimposed and interlocked. They also come in a huge variety of colors and shapes - a variety that is constantly increasing, as new products and packaging come onto the market every day.
As a result, this type of dataset presents a major problem of under-representation of certain classes. It is also extremely time-consuming and technical to annotate, so much so that it requires the supervision of waste industry experts.
For all these reasons, these labeled images are invaluable, either for improving an existing model, or for testing new, targeted ideas.
How is it innovative?
The main aim of this article is to provide a dataset for researchers to tackle the problems of sorting centers and develop new algorithms to overcome these challenges.
In their article, they publish:
- a fully labeled dataset ;
- an unlabeled dataset for semi-supervised learning;
- a dataset augmented with objects from TACO (Trash Annotation in COntext) to combat class imbalance.
The article also provides a baseline for instance segmentation using Mask-RCNN, TridentNet and DeepLabV3+.
What are its limits?
Image capture
Their models struggle to achieve the performance we see in the usual articles on object detection or instance segmentation. They also fall short of the performance we see on a similar conveyor. One possible explanation is the image capture protocol they use.
At Lixo, we use a camera closer to the conveyor with brighter light to reduce blur and improve image quality. We believe that the blur removal algorithm (SRN-Deblur) and the fish-eye removal they had to implement to get their final images diminish image quality. These transformations are the reason for the model's reduced performance on small objects.
.png)
Taxonomy
The second major limitation is the class taxonomy used. From the point of view of a waste management professional, it would be far too limited: there aren't enough classes to accurately classify a waste stream. At Lixo, we have a much more detailed class taxonomy. For example, we distinguish each plastic according to its resin (clear PET, colored PET, HDPE, etc.) to get a real idea of waste quality and recovery value.
Conclusion
In conclusion, we think it's great that university researchers are interested in waste management, and we're very excited to see what they can bring to the field.
But their approach also shows how important it is to keep in step with customer needs. Indeed, customers in the waste management sector don't just want to identify "contaminants", they also want to understand the share of contaminants in a certain stream. This share, whether given as a % of total objects or as a % of mass, requires the detection of 100% of the objects in the image.
Lixo shines in providing both (contaminant identification + target material) to its customers, and we believe this makes a huge difference in waste stream analysis!
Find out more:
- ZeroWaste dataset: Towards Deformable Object Segmentation in Cluttered Scenes, Dina Bashkirova, Mohamed Abdelfattah, Ziliang Zhu, James Akl, Fadi Alladkani, Ping HuVitaly Ablavsky, Berk Calli, Sarah Adel Bargal and Kate Saenko (2021)
