Impact of Noise in Large Real-World Datasets on Semi-Supervised Object Detection: A Case Study of Homeless Encampments Detection

Large image datasets have driven innovations in image-based machine learning, especially supervised object detection. However, manual annotation of such datasets has been challenging due to the issues related to labeling time, the subjectivity of human perception, and the scalability of dataset size. To that end, semi-supervised object detection (SSOD) methods have emerged that leverage a small amount of manually labeled and much larger unlabeled data. However, existing SSOD methods have been trained over clean, noise-free datasets such as MS-COCO and PASCAL VOC. They thus may fail to capture the reality, i.e., the heterogeneity and nuances of objects in the real world, which contain noises such as occlusion, illumination, and irregularity in the object’s shape. This study conducts an empirical analysis of the performance of semi-supervised object detection on a real-world custom dataset of homeless encampments. Results show that certain noise conditions may be ignored during the manual annotation process without causing a significant drop-off in SSOD performance across various metrics.