Abstract
DCF-ASN introduces a "coarse-to-fine" framework that marries the high-speed adaptability of Discriminative Correlation Filters (DCF) with the precise localization of Siamese networks. By using an online-updating DCF to roughly find the target and an offline-trained Attentional Siamese Network (ASN) to refine the result, the tracker achieves state-of-the-art accuracy at a real-time speed of 38 FPS. This hybrid approach effectively addresses the common "drift" problem in visual tracking while maintaining computational efficiency.
Introduction
Visual tracking—following an arbitrary object through a video sequence—is plagued by challenges like fast motion, occlusions, and background clutter. Currently, the field is split between two dominant approaches: Discriminative Correlation Filters (DCF), which are fast and adaptive but prone to drift, and Siamese Networks, which offer powerful similarity metrics but can be computationally heavy or fail during significant appearance changes if not updated online. DCF-ASN addresses the accuracy-speed trade-off by proposing a hybrid architecture that uses the strengths of one to mitigate the weaknesses of the other.
Methods
The core of the DCF-ASN framework operates in two distinct stages:
Coarse Tracking (Online DCF)
The system first uses a standard DCF module to predict the target's rough position and scale. This module is updated online using a learning rate to adapt to the target's changing appearance. It leverages Fast Fourier Transforms (FFT) to maintain high efficiency, acting as a "guide" that significantly narrows down the search space.
Fine Tracking (Offline ASN)
Once the DCF provides a coarse location, the system generates several candidate proposals around that area. These are processed by the Attentional Siamese Network (ASN):
- Attention Subnet: This component analyzes the initial ground-truth template to learn specific channel weights, identifying which features (e.g., color, edges, textures) are most important for the specific target.
- Estimation Subnet: It applies these learned weights to the search region and predicts a confidence score for each proposal. The proposal with the highest score becomes the final tracking result.
The entire system uses a ResNet-50 backbone and is trained using a Generalized Intersection over Union (GIoU) loss to improve bounding box regression.
Experimental Results
The authors evaluated DCF-ASN on seven major tracking benchmarks, including OTB100, VOT2018, LaSOT, GOT-10k, TrackingNet, and aerial benchmarks like UAV123 and UAVDT.
Key Performance Metrics
- Accuracy: On the GOT-10k dataset, it reached an Average Overlap (AO) of 0.612, outperforming established models like ATOM and DiMP.
- Speed: The tracker runs at approximately 38 FPS on a single NVIDIA 1080ti GPU, exceeding the real-time threshold.
- Efficiency: The model size is roughly 211MB, making it leaner and more suitable for deployment than competitors like DiMP (~364MB).
Robustness and Drift Mitigation
Qualitative results demonstrate the tracker's ability to relocate targets even after they are momentarily "lost" by the DCF module. This proves the effectiveness of the ASN refinement stage in correcting accumulated errors.
Discussion and Future Work
DCF-ASN successfully avoids the "online update trap" by keeping the complex Siamese localization module offline while updating the lightweight DCF online. This ensures adaptability without the high risk of the entire network forgetting the target.
While highly effective, future improvements could involve incorporating "anchor-free" object detection techniques to further enhance bounding box precision. Additionally, computational optimization techniques like tensor decomposition could enable the model to run on lower-power mobile or embedded devices.
Summary updated