In this work, we adopt metamorphic testing to evaluate the robustness of hand pose estimation on four state-of-the-art models: MediaPipe hands, OpenPose, BodyHands, and NSRM hand.
Occlusions, illumination variations and motion blur are indetified as the main obstacles to the performance of existing hand pose estimation models. Considering their influence on the HPE models, we transform the source test case obtianed from two public hand pose datasets: FreiHand and CMU Panoptic Hand to construct the corresponding follow-up test cases, and propose the following metamorphic relations:
The experimental results are uploaded and placed at the corresponding folders of this repositories at: Source test cases, MR1, MR2, MR3, and MR4.
Here are samples of hands in different test cases:
The code is released under the MIT license.