Related Work

Active Learning

Active learning aims to reduce annotation cost by selecting only the most informative unlabeled samples for human labeling. Common strategies include uncertainty sampling, diversity-based sampling, representativeness-based sampling, and query-by-committee methods.

These methods have been widely studied for classification tasks, where each selected sample usually requires only a single class label. However, object detection introduces a more complex annotation process because each image may contain multiple objects that need to be localized.

Active Learning for Object Detection

Active learning for object detection extends these ideas to detectors that predict both object categories and bounding boxes. Prior methods have explored uncertainty estimation, localization consistency, diversity constraints, core-set selection, and plug-and-play sampling strategies for selecting informative images.

Despite this progress, most existing methods mainly focus on which images should be queried. After selection, they still assume that human annotators will draw bounding boxes manually, which leaves a major part of the annotation cost unresolved.

Foundation Models and Binary Feedback

Recent vision and vision-language foundation models provide strong zero-shot abilities in image understanding, localization, and representation learning. These models create new opportunities for reducing annotation effort in object detection pipelines.

Instead of relying on a single foundation model to replace human annotators, CoMODAL uses multiple foundation models cooperatively. They help mine negative images, select informative queries, and generate candidate bounding boxes that humans can verify with binary accept/reject feedback.