Fast Visual Object Tracking with Rotated Bounding Boxes (1907.03892v5)

Abstract: In this paper, we demonstrate a novel algorithm that uses ellipse fitting to estimate the bounding box rotation angle and size with the segmentation(mask) on the target for online and real-time visual object tracking. Our method, SiamMask_E, improves the bounding box fitting procedure of the state-of-the-art object tracking algorithm SiamMask and still retains a fast-tracking frame rate (80 fps) on a system equipped with GPU (GeForce GTX 1080 Ti or higher). We tested our approach on the visual object tracking datasets (VOT2016, VOT2018, and VOT2019) that were labeled with rotated bounding boxes. By comparing with the original SiamMask, we achieved an improved Accuracy of 0.652 and 0.309 EAO on VOT2019, which is 0.056 and 0.026 higher than the original SiamMask. The implementation is available on GitHub: https://github.com/baoxinchen/siammask_e.

• The paper introduces SiamMask_E, a novel algorithm that uses ellipse fitting on segmentation masks to estimate accurate rotated bounding boxes for object tracking.
• SiamMask_E significantly improves tracking accuracy (e.g., 65.2% on VOT2019) while maintaining real-time performance at 80 fps, outperforming prior methods.
• This research provides a crucial step towards robust real-time tracking in complex environments, with implications for applications in robotics, surveillance, and autonomous systems.

Fast Visual Object Tracking with Rotated Bounding Boxes: An Overview

In this paper, the authors present a notable advancement in the domain of visual object tracking by introducing a novel algorithm that incorporates ellipse fitting to effectively estimate rotated bounding boxes for tracking objects in real-time. The proposed method, dubbed SiamMask_E, enhances the bounding box fitting mechanism of the existing SiamMask tracking algorithm, achieving improved accuracy while maintaining a high frame rate of 80 fps on standard GPU hardware.

Key Contributions and Methodology

The paper outlines several core contributions:

1. Enhanced Tracking Accuracy: By integrating ellipse fitting into the bounding box estimation process, SiamMask_E demonstrates a significant improvement in tracking accuracy. On the VOT2019 dataset, the authors report an accuracy of 65.2% and an Expected Average Overlap (EAO) of 30.9%, marking improvements of 5.6% and 2.6% over the original SiamMask algorithm.
2. Novel Algorithm for Rotated Bounding Boxes: The paper introduces a computationally efficient algorithm for estimating rotated bounding boxes from segmentation masks, leveraging the principles of ellipse fitting. This method involves calculating the rotation angle and scale of the bounding box by fitting an ellipse to the mask, thus facilitating more precise orientation and tighter bounding boxes.
3. Real-time Performance: The proposed algorithm retains the high-speed tracking capability of the original SiamMask, making it suitable for real-time applications such as surveillance, autonomous vehicles, and robotics, which demand both accuracy and efficiency.

The authors detail an approach that revolves around estimating the rotation angle of objects by fitting an ellipse to the segmentation mask generated by SiamMask. The algorithm is composed of two main steps: rotation angle estimation and scale calculation. The use of ellipse fitting, inspired by Fitzgibbon et al.'s conic fitting methodology, allows for refined computation of both the orientation and size of the bounding boxes.

Evaluation and Results

The paper conducts extensive evaluations on several benchmark datasets—VOT2016, VOT2018, and VOT2019—all of which feature rotated bounding box annotations. SiamMask_E outperforms several state-of-the-art Siamese network-based trackers, showcasing superior performance metrics in terms of accuracy and EAO.

• On the VOT2018 dataset, SiamMask_E achieves an accuracy of 62.7% and an EAO of 42.7%, surpassing other leading trackers, including SiamRPN++.
• Qualitative results highlight improved alignment with ground truth annotations, indicating that SiamMask_E effectively handles complex scenarios involving object rotation and occlusion.

Implications and Future Work

The advancement presented in this paper holds significant implications for the field of computer vision and real-time object tracking. By bridging the gap between accuracy and computational efficiency, SiamMask_E sets a new benchmark for applications requiring robust tracking capabilities. This development can be particularly beneficial in fields such as robotics, where real-time processing is paramount.

Looking forward, the authors suggest potential avenues for further research, including the integration of motion models to differentiate between camera and object motion, as well as accounting for dynamic distractors in tracking scenarios. Enhancing motion estimation could lead to even greater tracking precision and robustness in dynamic environments.

In conclusion, the paper provides a valuable contribution to the ongoing discourse in visual object tracking, proposing an innovative approach that marries elliptical geometry with real-time computing to attain advanced tracking accuracy. This work not only elevates the performance of current tracking systems but also paves the way for future explorations in the domain.