Small-Object-Oriented Pose Estimation With Structural Similarity Constraint

Le Han, Lei Zhao, Han Zhang, Zhiying Song, Pengfei Wang, Nenggan Zheng

Abstract: Remote measurement of animal pose and motion is essential in neuroscience and robotics. However, subjects like rats often appear as small objects in open environments, occupying a minimal field of view. The scarcity of visual information and high susceptibility to errors present significant challenges for remote measurement of the pose of these objects. This study introduces SalPose, a pose estimation model designed for small objects, addressing these challenges through rapid perception of small objects and keypoints’ structural information mining. First, we enhance the perception of weak visual features by integrating saliency detection and constructing pseudosaliency maps with annotated keypoints. Second, a scale-invariant structural similarity constraint is developed to facilitate the learning of intrinsic keypoint features and reduce the impact of small object size. In addition, to address the crucial gap in small-sized animal behavior analysis datasets in open environments, we introduce RP-5.7K, a rat pose dataset comprising exclusively small object data from the aerial perspective. Extensive experiments on the proposed dataset and two public datasets demonstrate the superiority of the proposed model. This study presents a novel method for remote motion measurement using visual information, mitigating the high sensitivity of small objects to distance-based measurement errors. The contributed dataset is valuable for visual measurement tasks involving small objects and animal behavior analysis. The RP-5.7K dataset is available at https://github.com/CSDLLab/RP-5.7K