Domain Adaptive Object Detection using Semi-Supervised Learning and Pseudo labeling based Self-Training

Abstract

In the area of computer vision object detection is concerned with detecting and localizing objects within images. Accurate localization of objects in an image plays a vital role in various computer vision tasks, such as object detection, tracking, and segmentation. However, when it comes to training machine learning models for object detection, a significant amount of annotated data is typically required for supervised learning. A common challenge arises when these models, trained on one data domain, encounter a drop-in performance when tested on data from an unseen domain. The existence of domain discrepancies or shifts poses significant hurdles to the generalization ability of the deep learning models. Even though augmenting the training data with samples from the new domain can improve performance, the process of collecting annotations for this data is often time-consuming and labor-intensive. To address this issue, researchers have created deep transfer learning approaches, particularly deep domain adaptation techniques. These approaches aim to adapt the model's learned representations to effectively handle domain discrepancies, enabling the network to generalize well across different domains. This study presents a novel Deep Domain adaptation approach using semi-supervised learning approach with pseudo labeling based self-training to solve object detection problem. It presents an approach to improve domain adaptive object detection through the integration of several techniques such as progressive confidence thresholding, non-maximal suppression, consistency regularization, and image-level and instance-level alignment are applied to enhance the performance of object detection models when dealing with domain shifts. The proposed method increases the detection accuracy and robustness by effectively addressing the challenges posed by domain adaptation. The experimental result indicates a significant improvement in performance of the initial Faster R-CNN model, which was exclusively trained on the source domain, when evaluated across diverse domains. As a result, the proposed approach improves the overall performance of deep neural networks by enhancing their ability to generalize across domains.