%20(1).png)
The Context
To launch our Lixo Journal Club article series, we chose to introduce the ZeroWaste: Toward Automated Waste Recycling paper by Bashkirova and al. that was presented in the 2022 Conference on Computer Vision and Pattern recognition!
This paper deviates from the norm seeing as it doesn’t suggest an improvement for a model architecture or loss computation, but rather describes a new dataset for object detection and segmentation comprised of waste management facilities images.
Why did we choose the ZeroWaste dataset paper?
The ZeroWaste dataset is the first open-source dataset that provides images from sorting facilities, specifically a high-quality paper conveyor. These images are different from well-known datasets such as COCO or PASCAL VOC.
Indeed, the objects present in these images are highly cluttered, deformed and/or shredded. These objects also encapsulate a large variety of shapes and colours — a variety which in practice is constantly expanding seeing new products, and therefore packages, are being released everyday.
Consequentially, seeing as these images reflect real world data, the classes tend to be highly unbalanced. Also, due to the complexity of the labels and objects, these datasets can be very tedious to annotate and require a waste-sorting expert’s supervision to validate.
For all of the reason mentioned above, annotated images of waste are very valuable to a company like Lixo, either to improve an existing model or test out new targeted ideas.
How is the ZeroWaste dataset innovative ?
The main goal of this paper is to provide a dataset for researchers to tackle real-world problems and develop new algorithms to overcome the challenges faced in sorting facilities.
In their paper, they release :
- a fully annotated dataset ;
- a dataset without labels for semi-supervised learning ;
- an augmented dataset with objects from TACO (Trash Annotation in COntext) to balance out classes.
They also provide a baseline for instance segmentation using Mask-RCNN, TridentNet and DeepLabV3+.
What are its limitations ?
Image quality
Their models struggle to reach the average performance yielded from comparable papers on object detection and instance segmentation. The results reported from this study are also drastically lower than those yielded from our models on a similar dataset. One possible explanation for this drop in performance could be their image capturing protocol.
At Lixo, we install our cameras closer to the conveyor belt and use brighter light fixtures which in turn leads to clearer, over-all higher quality images. In order to make their images usable, Bashkirova and al. had to implement a blur removal algorithm (SRN-Deblur) and a fish eye removal which unfortunately impacted their image quality. These transformations are the reason why their models perform poorly on small objects.
.png)
Taxonomy or scheme of classification
The second major limitation of this dataset is the taxonomy used for their annotations. From the point of view of a waste management professional, the range of labels is too limited in order to correctly sort a stream of waste.At Lixo we have a much more detailed class taxonomy. For instance we distinguish each plastic based on their resin (clear PET, colored PET, HDPE, etc) to get a real sense of waste quality and recovery value.
Conclusion
In conclusion, we find it wonderful that academic researchers are interested in waste management and we are very excited to see what they can bring to the field. But their approach also shows how valuable domain expertise is in the creation of a granular and pertinent datasets. Indeed, clients in the waste management industry do not only wish to classify objects but also to understand the share of a given class within a stream. That share, whether given in % of total objects or % of mass, requires the detection of all objects present in an image.
Lixo shines by providing both (identification of contaminants + target material) to its clients, and we believe it makes a huge difference in the analysis of waste stream!
To learn 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).
- Discover why we launch our AI Journal Club
