Box Supervised

Box-supervised learning aims to perform complex image analysis tasks, such as instance segmentation and object detection, using only bounding box annotations, a significantly less labor-intensive approach than pixel-level labeling. Current research focuses on developing novel algorithms and model architectures that effectively leverage this limited supervision, often employing techniques like pseudo-label generation, contrastive learning, and attention mechanisms to improve segmentation accuracy and address challenges like shape distortion and out-of-distribution detection. This approach holds significant promise for reducing the annotation burden in computer vision, enabling the training of high-performing models on datasets where detailed annotations are scarce or expensive to obtain.