A New Target-specific Object Proposal Generation Method for Visual Tracking

Guanjun Guo, Hanzi Wang, Yan Yan, Hong-Yuan Mark Liao, Bo LiarXiv 2018 · OpenReview

Abstract

Object proposal generation methods have been widely applied to many computer vision tasks. However, existing object proposal generation methods often suffer from the problems of motion blur, low contrast, deformation, etc., when they are applied to video related tasks. In this paper, we propose an effective and highly accurate target-specific object proposal generation (TOPG) method, which takes full advantage of the context information of a video to alleviate these problems. Specifically, we propose to generate target-specific object proposals by integrating the information of two important objectness cues: colors and edges, which are complementary to each other for different challenging environments in the process of generating object proposals. As a result, the recall of the proposed TOPG method is significantly increased. Furthermore, we propose an object proposal ranking strategy to increase the rank accuracy of the generated object proposals. The proposed TOPG method has yielded significant recall gain (about 20%-60% higher) compared with several state-of-the-art object proposal methods on several challenging visual tracking datasets. Then, we apply the proposed TOPG method to the task of visual tracking and propose a TOPG-based tracker (called as TOPGT), where TOPG is used as a sample selection strategy to select a small number of high-quality target candidates from the generated object proposals. Since the object proposals generated by the proposed TOPG cover many hard negative samples and positive samples, these object proposals can not only be used for training an effective classifier, but also be used as target candidates for visual tracking. Experimental results show the superior performance of TOPGT for visual tracking compared with several other state-of-the-art visual trackers (about 3%-11% higher than the winner of the VOT2015 challenge in term of distance precision).

Introduction

Visual tracking is a fundamental task in computer vision, typically implemented via a "tracking-by-detection" paradigm. This approach involves sampling candidate regions (proposals) around the target's last known position and using a classifier to identify the most likely target location. However, generic object proposal methods—such as Selective Search or EdgeBoxes—are designed for general object detection and often fail in tracking scenarios characterized by motion blur, low contrast, or deformation.

This paper introduces a Target-specific Object Proposal Generation (TOPG) method that leverages video context and specific target appearance to overcome these challenges. By moving away from generic "objectness," TOPG focuses on finding a specific target, leading to higher recall and more robust tracking.

Methodology

Target-specific Object Proposal Generation (TOPG)

The core innovation of TOPG is the shift from generic objectness to target-specific cues. The method employs a two-pronged approach to ensure high-quality proposal generation:

  1. Tri-map and Color Cues: The algorithm partitions the initial frame into three regions: definite foreground (the target), definite background, and a transition region. Using non-parametric density estimation, it creates a response map where each pixel represents the probability of belonging to the target.
  2. Complementary Cue Integration: TOPG integrates color and edge information to maintain robustness. While color is resilient to motion blur, edges are more robust to illumination changes. By performing edge detection on both the raw image and the color-based response map, TOPG ensures high-quality proposals even under severe blur.
  3. Ranking Strategy: To improve accuracy, generated proposals are ranked based on their affinity to the original target in terms of shape (contour similarity), color (mean response map value), and size.
Visualization of target-specific object proposals for a car (top) and a cartoon character (bottom), demonstrating the high recall of the TOPG method.

The TOPGT Tracker

The TOPG method is integrated into a tracking framework called TOPGT, which utilizes a Convolutional Neural Network (CNN) architecture. Unlike standard trackers that rely on random Gaussian sampling for training, TOPGT uses the high-quality proposals generated by TOPG. This provides the network with "hard negatives"—distractors that closely resemble the target—making the final classifier significantly more discriminative and reducing the likelihood of tracker drift.

Experimental Results

The proposed method was evaluated on two major benchmarks: OTB-2013 (50 sequences with various challenges) and UAV20L (20 long-term high-definition sequences from drones).

  • Proposal Recall: TOPG achieved a recall gain of 20% to 60% higher than state-of-the-art generic methods on both datasets.
  • Tracking Accuracy: On OTB-2013, the TOPGT tracker achieved a distance precision of 93.1%, outperforming the VOT2015 challenge winner by approximately 3–11%.
  • Long-term Robustness: On the UAV20L dataset, TOPGT surpassed the well-known MDNet by 9% in precision and 7% in success rate, proving its effectiveness for long-duration tracking where targets frequently disappear and reappear.

Conclusion

TOPG demonstrates that incorporating target-specific context into the proposal generation phase significantly improves both recall and tracking accuracy. By providing high-quality candidates and hard negatives for CNN training, the TOPGT tracker sets a new standard for robust visual tracking in challenging environments. Future work may involve optimizing the computational efficiency of the proposal module to enable real-time performance on low-power hardware.

Summary updated

Abstract

Object proposal generation methods have been widely applied to many computer vision tasks. However, existing object proposal generation methods often suffer from the problems of motion blur, low contrast, deformation, etc., when they are applied to video related tasks. In this paper, we propose an effective and highly accurate target-specific object proposal generation (TOPG) method, which takes full advantage of the context information of a video to alleviate these problems. Specifically, we propose to generate target-specific object proposals by integrating the information of two important objectness cues: colors and edges, which are complementary to each other for different challenging environments in the process of generating object proposals. As a result, the recall of the proposed TOPG method is significantly increased. Furthermore, we propose an object proposal ranking strategy to increase the rank accuracy of the generated object proposals. The proposed TOPG method has yielded significant recall gain (about 20%-60% higher) compared with several state-of-the-art object proposal methods on several challenging visual tracking datasets. Then, we apply the proposed TOPG method to the task of visual tracking and propose a TOPG-based tracker (called as TOPGT), where TOPG is used as a sample selection strategy to select a small number of high-quality target candidates from the generated object proposals. Since the object proposals generated by the proposed TOPG cover many hard negative samples and positive samples, these object proposals can not only be used for training an effective classifier, but also be used as target candidates for visual tracking. Experimental results show the superior performance of TOPGT for visual tracking compared with several other state-of-the-art visual trackers (about 3%-11% higher than the winner of the VOT2015 challenge in term of distance precision).
BibTeX
@misc{guo2018new,
  author = {Guanjun Guo and Hanzi Wang and Yan Yan and Hong-Yuan Mark Liao and Bo Li},
  title = {{A New Target-specific Object Proposal Generation Method for Visual Tracking}},
  year = {2018},
  eprint = {1803.10098},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1803.10098},
}
Figure