Empirical Study of Real-time One-Stage Object Detection Methods on Recyclable Waste Dataset

One-stage object detection methods have proven their advantage in terms of both speed and accuracy for addressing vision tasks in real-time scenarios, including Recyclable Waste detection, which has become a prevalent topic during the COVID-19 pandemic. Previous research into this subject has faced many obstacles, mainly due to the requirement of detecting highly deformable and often translucent objects in cluttered scenes without the context information usually present in human-centric datasets. In this paper, we aim to explore the performance of state-of-the-art one-stage object detectors on ZeroWaste dataset, the first in-the-wild industrial-grade waste detection benchmark. Our experiments showed that recent one-stage detectors, namely the YOLO-based detectors, can obtain very competitive results on the benchmark. YOLOv7, thanks to its many improvements, is the current best performer at 33.2% mAP on the ZeroWaste benchmark, to the best of our knowledge. Implementation details are available at our GitHub repository.