Journal article
Surrogate Model-based Explainability Method for 3D Object Tracking
IEEE transactions on multimedia, Early Access
2026
Abstract
3D object tracking is a popular research topic in computer vision. However, it remains a challenging task to improve the trustworthiness of deep 3D trackers, due to the complex network structures of these black-box models. In this paper, an explainability method is designed to generate local explanations for 3D object tracking, which trains an interpretable surrogate model to reveal the contributions of superpoints in the search area to the decision-making of a 3D tracker. Specifically, local points within the search area with comparable geometric features are aggregated into superpoints, which serve as the basis for the explanation, and non-negative matrix factorization is employed to uncover latent features of the search area. In contrast to the commonly used voxels, the representation of superpoints captures rich semantic information of the search area, and facilitates an intuitive understanding of the explanations. To train the surrogate model, an adaptive downsampling strategy is proposed to generate the sample set by removing superpoints, with the latent contributions of superpoints to predictions considered to improve explanation efficacy. Additionally, each sample is weighted during training to better approximate the decision of a 3D tracker. Experiments have demonstrated that the proposed explainability approach can effectively provide explanations to multiple deep tracking models on three large-scale datasets.
Details
- Title
- Surrogate Model-based Explainability Method for 3D Object Tracking
- Authors/Creators
- Riran Cheng - University of Electronic Science and Technology of ChinaXupeng Wang - University of Electronic Science and Technology of ChinaFerdous Sohel - Murdoch UniversityYu Peng - University of Electronic Science and Technology of ChinaHang Lei - University of Electronic Science and Technology of China
- Publication Details
- IEEE transactions on multimedia, Early Access
- Publisher
- IEEE
- Number of pages
- 12
- Identifiers
- 991005904077607891
- Copyright
- © Copyright 2026 IEEE - All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies.
- Murdoch Affiliation
- School of Information Technology
- Language
- English
- Resource Type
- Journal article
Metrics
1 File views/ downloads
4 Record Views