EasyClick: A New Web-based Image Annotation Tool for Person Re-identification Using Deep Learning
Md. Khayrul Bashar, Yusei Fujimoto, Yuichiro Kaneoka · 2024
Image annotation is a vital step for model building and object recognition. Although fully automatic annotation is expected, it still has limitations in the scenario like person re-identification (ReID) where multi-camera images are involved. This difficulty arises due to complex intra-class variations in illumination, pose, viewpoint, blur, low resolution, and occlusion. Researchers typically use manual annotation tools that annotate gallery images by searching them one-by-one for each query image, which is a tedious and time-consuming task. In the study, we propose a new web-based semi-automatic annotation tool, called “EasyClick”, which capitalizes the capacity of a deep learning model, called omni-scale network (OSNet), with cosine similarity metric and a clustering algorithm. Our proposed approach is a versatile one that can provide two ranking suggestions with and without using a hierarchical clustering algorithm. Given a query image from one camera, this tool can retrieve a small subset of the most similar images or a few clusters of ranked images from another camera. Users can then select all relevant images to a query by easily clicking the displayed images. Several experiments with two datasets having 43,246 (802 persons) and 594 (124 persons) multi-camera images showed promising performance of the proposed tool in terms of speed and accuracy when compared to the popular CVAT annotation tool.