Depth Attention for Robust RGB Tracking

TL;DR

This work proposes a new framework that leverages monocular depth estimation to counter the challenges of tracking targets that are out of view or affected by motion blur in RGB video sequences and forms a simple framework that allows seamlessly integration of depth information with state of the art tracking algorithms, without RGB-D cameras.

Abstract

RGB video object tracking is a fundamental task in computer vision. Its effectiveness can be improved using depth information, particularly for handling motion-blurred target. However, depth information is often missing in commonly used tracking benchmarks. In this work, we propose a new framework that leverages monocular depth estimation to counter the challenges of tracking targets that are out of view or affected by motion blur in RGB video sequences. Specifically, our work introduces following contributions. To the best of our knowledge, we are the first to propose a depth attention mechanism and to formulate a simple framework that allows seamlessly integration of depth information with state of the art tracking algorithms, without RGB-D cameras, elevating accuracy and robustness. We provide extensive experiments on six challenging tracking benchmarks. Our results demonstrate that our approach provides consistent gains over several strong baselines and achieves new SOTA performance. We believe that our method will open up new possibilities for more sophisticated VOT solutions in real-world scenarios. Our code and models are publicly released: https://github.com/LiuYuML/Depth-Attention.

Figures (8)

• Figure 1: Comparison of our approach (depth attention) with RGB-D Tracking.
• Figure 2: Proposed depth attention to improve RGB tracking.
• Figure 3: This figure shows the original images, their depth estimations, $Z_K$ values, I$_{dc}$, and tracking outcomes across four different sequences. Noticeably, when the baseline tracker starts to drift, our proposed method effectively prevents this by generating a mask with zero values in the background area. The proposed depth attention based masking enhances the overall tracking performance.
• Figure 4: We compare the original algorithm, the depth attention module with a fixed hyperparameter (Th=1.5, $k_1=0.02$), and the depth attention module with an adaptive hyperparameter on $k_1$. We elaborate on the base algorithm, dataset, and the metrics used along the axes.
• Figure 5: Visual comparison between KeepTrack and KeepTrack+DA (ours). It can be seen that, our method is robust under the challenging scenarios of occlusion and motion blur.