Ocean: Object-aware Anchor-free Tracking (2006.10721v2)

Abstract: Anchor-based Siamese trackers have achieved remarkable advancements in accuracy, yet the further improvement is restricted by the lagged tracking robustness. We find the underlying reason is that the regression network in anchor-based methods is only trained on the positive anchor boxes (i.e., $IoU \geq0.6$). This mechanism makes it difficult to refine the anchors whose overlap with the target objects are small. In this paper, we propose a novel object-aware anchor-free network to address this issue. First, instead of refining the reference anchor boxes, we directly predict the position and scale of target objects in an anchor-free fashion. Since each pixel in groundtruth boxes is well trained, the tracker is capable of rectifying inexact predictions of target objects during inference. Second, we introduce a feature alignment module to learn an object-aware feature from predicted bounding boxes. The object-aware feature can further contribute to the classification of target objects and background. Moreover, we present a novel tracking framework based on the anchor-free model. The experiments show that our anchor-free tracker achieves state-of-the-art performance on five benchmarks, including VOT-2018, VOT-2019, OTB-100, GOT-10k and LaSOT. The source code is available at https://github.com/researchmm/TracKit.

• The paper introduces a novel anchor-free method that directly predicts bounding boxes from target pixels to enhance tracking accuracy.
• It employs a dual-network design utilizing regression and classification with object-aware feature alignment to distinguish foreground from background.
• It achieves state-of-the-art results with a 0.467 EAO on VOT-2018 and operates at up to 58 FPS, ensuring robust real-time performance.

An Analytical Overview of "Ocean: Object-aware Anchor-free Tracking"

This paper introduces a novel approach to visual object tracking by presenting an object-aware anchor-free framework named "Ocean." The authors focus on addressing limitations inherent in anchor-based Siamese networks, specifically tackling challenges related to tracking robustness. They propose an innovative network architecture, which directly predicts the position and scale of target objects in an anchor-free manner, thus enabling the tracker to rectify imprecise predictions more effectively.

Methodology

The Ocean framework consists of two main components: the regression network and the classification network. The regression network departs from traditional anchor-based methods, aiming to predict boundary box extents directly from each pixel within the target object. This approach allows each pixel in the groundtruth box to contribute during training, which enhances prediction rectification capabilities.

The classification network operates on an object-aware feature, resolved through a feature alignment module. This feature is pivotal in distinguishing foreground from background more reliably, owing to the adaptive feature transformation module incorporated into the framework. By sampling features directly from predicted bounding boxes, the model can maintain robustness against variations in object scale and position.

Results and Evaluation

The proposed Ocean tracker demonstrates state-of-the-art performance across five benchmark datasets: VOT-2018, VOT-2019, OTB-100, GOT-10k, and LaSOT. Notably, in the VOT-2018 benchmark, the tracker achieves an Expected Average Overlap (EAO) of 0.467, outperforming leading contemporaries like SiamRPN++ by significant margins. Moreover, it sustains efficient frame rates, achieving up to 58 FPS, affirming its real-time tracking capability.

Implications

The implications of adopting an anchor-free architecture are multifaceted. Practically, it enhances tracking robustness in scenarios where target objects undergo rapid motions or significant occlusions. Theoretically, it opens avenues for further research into anchor-independent methodologies, potentially influencing the design of future tracking systems.

Future Directions

Building upon its demonstrated efficacy, further developments in the Ocean framework could involve refining the online update mechanism to adapt more dynamically to changing target appearances. Researchers may also explore applying similar anchor-free strategies to broader domains such as video object detection and segmentation.

Conclusion

The Ocean framework captures a significant stride in object tracking, offering a promising alternative to anchor-based systems. Its emphasis on direct prediction and feature alignment encapsulates a forward-thinking approach, contributing substantial advancements to the visual tracking field.