Cross-Pollination-of-Knowledge-for-Object-Detection-in-Domain-Adaptation-for-Industrial-Automation

Artificial Intelligence (AI) transforms industries by automating activities and enhancing efficiency using real-time Object Detection (OD) applications. The OD with modern computer vision methods empowers systems to automate processes, analyze challenging visual data, and make data-driven choices, eventually boosting productivity. Domain adaptation (DA) for OD has recently depicted ample attention due to its competence in identifying target objects with no annotations. Recently, new methodologies integrating traditional cross-disciplinary domain modeling with advanced deep learning techniques are increasingly recognized as crucial for solving complex AI and real-time problems. This work proposes a simple but effective inter-disciplinary Cross-Pollination of knowledge (CPK) strategy for source and target domains in Domain Adaptation. The CPK originates from the Botany domain; it fuses the target samples in the source samples at the input level, and the fusion of a random and unique number of target samples supports the detector in feature alignment and generalization with the source domain. This work also introduced the newly constructed digit recognition dataset (Planeat), which comprises 231 images. For the fair comparison, this study employs an Unsupervised Domain Adaptation (UDA) method that takes advantage of unsupervision to train the target and source domains synchronously. UDA method uses target data, extracts the high-confidence region, crop and applies augmentation techniques to these regions, and adapts UDA for OD. The proposed CPK performs efficiently through extensive experiments in five different datasets under cross-weather, cross-camera, and synthetic-to-real cases, outperforming the existing UDA results by the margin of 10.9% of mean Average Precision (mAP).

Setup

For Self supervised domain adaptation experiments, follow

https://github.com/MohamedTEV/DACA

For supervised CPK experiments, follow YOLOv5

https://github.com/ultralytics/yolov5

Datasets

For Self supervised domain adaptation, downlaod and use the dataset given at https://github.com/MohamedTEV/DACA

For supervised CPK, Use Planeat and Mixture dataset given in the dataset section

Experiments setup

coming soon ...