Neural Network Keypoint Detection With Low-Resolution Heatmaps

The patent applies preliminary action by performing keypoint grouping and association directly in the low-resolution heatmap domain before upscaling. The neural network predicts keypoint locations and performs clustering operations on the downsampled feature maps, avoiding the need for computationally intensive post-processing operations like bicubic interpolation and vector field alignment on high-resolution images. This resolves the contradiction by achieving accurate grouping at lower computational cost.

The patent uses copying by working with compressed representations (heatmaps) of keypoint locations rather than operating directly on the full-resolution image data. The neural network generates heatmap maps that encode keypoint probability distributions, and all grouping operations are performed on these compact representations. This copying to a simplified domain reduces computational complexity while preserving the essential information needed for accurate keypoint association.

The patent applies partial action by predicting only the keypoints that are actually visible and detectable for each object instance, rather than forcing prediction of all predefined keypoints. The neural network learns to identify which keypoints are present and predicts only those, using attention mechanisms or masking to handle missing keypoints. This resolves the contradiction by achieving reliable skeleton detection without generating false positives from forced predictions of missing keypoints.

The patent uses parameter changes by dynamically adjusting the keypoint prediction process based on detected object characteristics. The system modifies which keypoints are predicted and how they are grouped based on confidence scores, object scale, and visibility conditions. This adaptive approach allows complete skeleton detection when possible while avoiding false positives when keypoints are occluded or missing.

The patent applies dimensionality change by transforming the keypoint detection problem from operating on high-resolution spatial dimensions to operating on lower-resolution heatmap dimensions. The neural network downsamples the feature maps and performs keypoint prediction in this compressed dimensional space, then uses the learned representations to accurately locate keypoints. This resolves the contradiction by achieving good accuracy through smart use of the heatmap dimension rather than brute-force high-resolution processing.

The patent applies local quality by concentrating computational resources on regions and features that are most important for keypoint detection. The neural network uses selective refinement, applying higher processing detail only to areas with high keypoint probability or ambiguous detections, while using coarser processing for clearly detected keypoints. This resolves the contradiction by optimizing the balance between accuracy and computational cost locally rather than uniformly across the entire image.