3D objects localization from a single calibrated viewpoint : a study of basketballs
(2024)
Files
Gabriel-VanZandycke_thesis.pdf
Open Access - Adobe PDF
- 29.82 MB
Details
- Authors
- Supervisors
- De Vleeschouwer, Christophe
- Abstract
- Localizing objects in 3D space is a fundamental computer vision task that finds applications in various domains such as autonomous driving and augmented reality. In the context of team sports analytics, the accurate localization of the ball provides fundamental cues of the game being played. Traditionally, this localization involves triangulation from multiple viewpoints. However, there are many practical cases where 3D localization is valuable even though only a single viewpoint is available. Moreover, multicamera systems are complex and expensive to deploy. In this thesis, we address the task of 3D object localization from a single calibrated viewpoint and apply it to basketballs. For this, we use a two stages approach. In the first stage, we leverage a deep neural network to detect basketballs in the image space. Based on a fast multi-scale segmentation model trained to predict the ball segmentation mask, our original method exhibits state of the art performances. Given the detections from our detector, for the second stage, we propose and compare two alternative neural-network-based approaches to estimate the ball 3D position. In the first approach, the neural network is trained to estimate ball diameter in the image space, and knowledge of real ball diameter is used to convert the prediction into depth. In the second approach, the neural network directly predicts ball height in the image space. It doesn't make any assumption on the object and could be extended to other domains beyond sport.
- Affiliations
Citations
Van Zandycke, G. (2024). 3D objects localization from a single calibrated viewpoint : a study of basketballs. https://hdl.handle.net/2078.5/214170