On Hyperbolic Embeddings in 2D Object Detection
Authors: Christopher Lang, Alexander Braun, Abhinav Valada
Summary: Object detection, for essentially the most half, has been formulated within the euclidean house, the place euclidean or spherical geodesic distances measure the similarity of a picture area to an object class prototype. On this work, we examine whether or not a hyperbolic geometry higher matches the underlying construction of the article classification house. We incorporate a hyperbolic classifier in two-stage, keypoint-based, and transformer-based object detection architectures and consider them on large-scale, long-tailed, and zero-shot object detection benchmarks. In our in depth experimental evaluations, we observe categorical class hierarchies rising within the construction of the classification house, leading to decrease classification errors and boosting the general object detection efficiency.
