TL;DR
This paper introduces an adaptive proposal selection algorithm designed to improve scale estimation in visual object tracking. By using color similarity and a confidence-weighted selection strategy to filter out redundant "EdgeBox" proposals, the authors created a tracker that is both more accurate and significantly faster than previous state-of-the-art methods.
What Problem Does This Paper Address?
Visual object tracking—following a specific object through a video—is plagued by the challenge of scale variation. As an object moves closer to or further from the camera, or rotates and deforms, its bounding box must resize accurately to maintain a "lock."
Traditional Correlation Filter (CF) trackers often struggle with this. Some try to solve it by testing a fixed set of scales (which is computationally expensive and inflexible), while others use "detection proposals" (like EdgeBoxes) to suggest potential bounding boxes. However, proposal generators often produce hundreds of redundant or low-quality candidates. These "distractors" not only slow down the tracker but can also lead to "drift," where the tracker accidentally starts following a piece of the background or a similar-looking object.
Methods
The authors propose a framework called CFAPS (Correlation Filter tracking with Adaptive Proposal Selection). The core innovation lies in how it whittles down the list of candidate boxes:
- Color Similarity Filtering: The algorithm represents the target and the proposals using color histograms in the HSV color space. It measures the similarity between proposals and two "instances": the very first frame (the "gold standard" appearance) and the most recent frame.
- Adaptive Selection Strategy: Not all frames are equally reliable. The authors track "contaminated frames"—frames where the tracker's confidence is low. If the recent tracking has been shaky, the algorithm automatically places more trust in the appearance of the object from the first frame to select the best proposals.
- Coarse-to-Fine Deep Features: To further boost performance, they integrated this selection logic with deep features from a pre-trained VGG-Net (specifically layers conv3-4, conv4-4, and conv5-4). This version, called DeepCFAPS, uses high-level semantic information to refine the tracking.
- Fine-Tuning: Once the best proposal is selected, a damping factor is used to smoothly update the target's state, preventing jittery box size changes.
Datasets
The researchers evaluated their approach on two major industry-standard benchmarks:
- OTB2013: A collection of 50 video sequences.
- OTB2015: An expanded version with 100 sequences, featuring challenging attributes like illumination changes, occlusion, and deformation.
Key Insights and Ideas
- Less is More: By discarding roughly half of the redundant proposals based on color similarity, the tracker actually becomes more accurate. This proves that many proposals generated by edge-based methods are actually "distractors" that lead the model astray.
- Temporal Reliability: The use of a "contaminated frame" counter ($\Delta_i$) is a clever way to handle model drift. It acknowledges that the tracker's internal model of the object can become corrupted over time, and provides a mathematical way to "re-anchor" the search to the original starting image.
- Efficiency Gains: Unlike many deep learning improvements that trade speed for accuracy, this method improves both. Filtering proposals reduces the number of expensive correlation operations required per frame.
Key Results
- Speed: The baseline KCFDPT tracker runs at 28.6 frames per second (fps). The proposed CFAPS tracker hits 40.1 fps—a nearly 40% speedup—while also being more accurate.
- Accuracy: On the OTB2015 dataset, CFAPS outperformed the baseline by 2.5% in Distance Precision (DP) and 1.6% in Area Under Curve (AUC).
- Scale Robustness: In specific tests for scale variation, the deep version (DeepCFAPS) outperformed the HCF tracker by a significant 5.9% on the OTB2013 dataset.
- Qualitative Success: Qualitative results showed that while baseline trackers often failed during fast motion or large deformations (like a gymnast twisting), CFAPS maintained a tight, accurately sized bounding box.
Limitations and Future Work
While the paper demonstrates strong results, a few areas remain for exploration:
- Color Dependency: Because the proposal selection relies heavily on HSV color histograms, the tracker might struggle in environments with very similar foreground/background colors (camouflage) or in grayscale video.
- Proposal Generation: The method still relies on EdgeBoxes as an external proposal generator. Future work could involve integrating the proposal generation more deeply into the neural network architecture itself.
- Computational Overhead of Deep Features: While the hand-crafted feature version (CFAPS) is very fast, the deep version (DeepCFAPS) requires a high-end GPU (NVIDIA GTX TITAN) to maintain performance, which may limit its use in mobile or embedded robotics.
Summary updated