Bottom-up multi-person pose estimation approaches use heatmaps with auxiliary predictions to estimate joint positions and affiliations simultaneously. Recently, various combinations of auxiliary predictions and heatmaps have been proposed to improve performance. The predictions are directly supervised by the corresponding L2 loss function. However, the lack of more explicit supervision results in low feature utilization and contradictions between predictions within a single model. To address these issues, this chapter proposes (1) a new loss function that employs self-supervised heatmaps to reduce prediction contradictions and spatial-sequential attention to enhance feature extraction by neural networks; (2) a new combination of predictions composed of heatmaps, part affinity fields (PAFs), and our block-inside offsets to correct pixel-level joint positions and further demonstrate the effectiveness of the proposed loss function. Experiments are conducted on the MS COCO keypoint dataset, using OpenPose as the baseline model. Our method outperforms the baseline overall. On the COCO verification dataset, the mAP of OpenPose trained with our proposals exceeds that of the baseline OpenPose by 5.5%.