Robust Object Detection via Kronecker Tensor Decomposition: Theory, Algorithms, and Applications

Salman Ahmadi-Asl, Roman Garaev, Hamidreza Behjoo, Asad Masood Khattak, Manuel Mazzara

Keywords: Kronecker tensor decompositions, randomized algorithms, object detection

Abstract: Object detection in real-world scenarios faces significant challenges, including occlusions, sensor noise, limited training data, and computational constraints, particularly in video analysis. This paper presents a comprehensive framework leveraging tensor completion methods to address these fundamental limitations. We demonstrate that tensor completion provides a mathematical foundation for handling missing and corrupted data by exploiting the multi-dimensional structure inherent in visual information. Our approach encompasses three key contributions: (1) a theoretical analysis of tensor rank properties in visual data, (2) novel algorithms integrating tensor completion with modern detection pipelines, and (3) extensive experimental validation across multiple benchmarks. Results show that our methods improve detection accuracy under severe occlusion, reduce computational cost in video detection while maintaining accuracy, and enhance robustness to various types of corruption. The proposed framework establishes tensor completion as a fundamental tool for building more robust and efficient object detection systems.

Submission Number: 20