This paper introduces associative embedding, a novel method for end-to-end learning of joint detection and grouping in computer vision tasks. The method allows a network to simultaneously output detections and group assignments, enabling efficient integration into existing architectures that produce pixel-wise predictions. The core idea is to assign a real number (tag) to each detection, which identifies the group it belongs to. These tags are used to group detections into meaningful structures, such as people in multi-person pose estimation or object instances in instance segmentation. The method is applied to multi-person pose estimation and instance segmentation, achieving state-of-the-art results on the MPII and MS-COCO datasets. In multi-person pose estimation, the network detects body joints and groups them into individuals using associative embedding tags. In instance segmentation, the network detects pixels and groups them into object instances based on similar tags. The approach is simple, generic, and can be applied to various vision tasks that involve detection and grouping. The method is trained using a loss function that encourages similar tags for detections from the same group and different tags for detections from different groups. This loss is applied to candidate detections that match the ground truth. The network is trained end-to-end, without requiring separate stages for detection and grouping. The approach is integrated with a stacked hourglass architecture, which is effective for dense pixel-wise prediction. The method is evaluated on two datasets: MS-COCO and MPII Human Pose. On MS-COCO, the method achieves state-of-the-art performance in multi-person pose estimation. On MPII, it also achieves state-of-the-art results. The method is also applied to instance segmentation, where it achieves reasonable results. The approach is shown to be effective in both tasks, demonstrating its versatility and effectiveness in computer vision. The method is general enough to be applied to other vision problems, such as multi-object tracking in video. The associative embedding loss can be implemented in any network that produces pixel-wise predictions, making it easy to integrate with other state-of-the-art architectures.This paper introduces associative embedding, a novel method for end-to-end learning of joint detection and grouping in computer vision tasks. The method allows a network to simultaneously output detections and group assignments, enabling efficient integration into existing architectures that produce pixel-wise predictions. The core idea is to assign a real number (tag) to each detection, which identifies the group it belongs to. These tags are used to group detections into meaningful structures, such as people in multi-person pose estimation or object instances in instance segmentation. The method is applied to multi-person pose estimation and instance segmentation, achieving state-of-the-art results on the MPII and MS-COCO datasets. In multi-person pose estimation, the network detects body joints and groups them into individuals using associative embedding tags. In instance segmentation, the network detects pixels and groups them into object instances based on similar tags. The approach is simple, generic, and can be applied to various vision tasks that involve detection and grouping. The method is trained using a loss function that encourages similar tags for detections from the same group and different tags for detections from different groups. This loss is applied to candidate detections that match the ground truth. The network is trained end-to-end, without requiring separate stages for detection and grouping. The approach is integrated with a stacked hourglass architecture, which is effective for dense pixel-wise prediction. The method is evaluated on two datasets: MS-COCO and MPII Human Pose. On MS-COCO, the method achieves state-of-the-art performance in multi-person pose estimation. On MPII, it also achieves state-of-the-art results. The method is also applied to instance segmentation, where it achieves reasonable results. The approach is shown to be effective in both tasks, demonstrating its versatility and effectiveness in computer vision. The method is general enough to be applied to other vision problems, such as multi-object tracking in video. The associative embedding loss can be implemented in any network that produces pixel-wise predictions, making it easy to integrate with other state-of-the-art architectures.