Descriptive Attributes for Language-Based Object Keypoint Detection

Multimodal vision and language (VL) models have recently shown strong performance in phrase grounding and object detection for both zero-shot and finetuned cases. We adapt a VL model (GLIP) for keypoint detection and evaluate on NABirds keypoints. Our language-based keypoints-as-objects detector GLIP-KP outperforms baseline topdown keypoint detection models based on heatmaps and allows for zero-and few-shot evaluation. When fully trained, enhancing the keypoint names with descriptive attributes gives a significant performance boost, raising AP by as much as 6.0, compared to models without attribute information. Our model exceeds heatmap-based HRNet's AP by 4.4 overall and 8.4 on keypoints with attributes. With limited data, attributes raise zero-/one-/few-shot test AP by 1.0/3.4/1.6, respectively, on keypoints with attributes.