AWS Builder Center

Evaluating Cross-Dataset Generalisation of Computer Vision Models for Construction and Demolition Waste Recognition

Nowadays, construction and demolition wastes has become one of the biggest threat to the world. The CDW management deadly play a crucial rule to preserve recyclability, reusability of the waste material. This study not underline on the type of deep learning models that used for waste recognition but evaluate the performance of the model when there are a domain shift

Construction and Demolition Waste (CDW) is one of the largest global waste streams contributors. The effective management and recovery of the waste material highly depends on the reliable identification and separation of materials. Computer vision and deep learning have been widely applied to automate CDW recognition to replace manual sorting methods by applying benchmark datasets and object-detection models reporting high accuracy. However, the high accuracy is the result of both training and testing set is derived from the same dataset whereas CDW datasets differ substantially in acquisition conditions, viewpoint, lighting, clutter and material taxonomy. These differences lead to domain shift as performance can decline when a model is applied to an independently collected dataset. Hence, a strong within-dataset accuracy may not indicate real-world reliability.However, it remains unclear how much performance declines across datasets, which material classes are most affected, and whether limited target-domain fine-tuning can reduce this decline. In order to address the research gap, this study proposes a quantitative, experimental evaluation of the cross-dataset generalisation of a pretrained YOLO-based object detector for CDW recognition. The model is trained on the CODD dataset as source which contains a set of aligned waste material classes and assessed on the independently gathered CDW-Seg dataset as target. There are three experimental conditions, a within-dataset baseline, a cross-dataset evaluation and an evaluation on small-sample target-domain fine-tuning. The evaluation of the model performance is assessed by using mAP@0.5, mAP@0.5:0.95, precision, recall, F1-score and class-level recall while supported by class-level and error analysis. The study is expected to quantify the degradation of the model when there is a domain shift and identify the classes and visual conditions that drive it. The research studies how far limited target data recovers the lost performance, providing evidence on the transferability of CDW recognition models before deployment.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article