Delving into the Trajectory Long-tail Distribution for Muti-object Tracking

Sijia Chen,En Yu,Jinyang Li,Wenbing Tao

2024-05-24

Abstract:Multiple Object Tracking (MOT) is a critical area within computer vision, with a broad spectrum of practical implementations. Current research has primarily focused on the development of tracking algorithms and enhancement of post-processing techniques. Yet, there has been a lack of thorough examination concerning the nature of tracking data it self. In this study, we pioneer an exploration into the distribution patterns of tracking data and identify a pronounced long-tail distribution issue within existing MOT datasets. We note a significant imbalance in the distribution of trajectory lengths across different pedestrians, a phenomenon we refer to as ``pedestrians trajectory long-tail distribution''. Addressing this challenge, we introduce a bespoke strategy designed to mitigate the effects of this skewed distribution. Specifically, we propose two data augmentation strategies, including Stationary Camera View Data Augmentation (SVA) and Dynamic Camera View Data Augmentation (DVA) , designed for viewpoint states and the Group Softmax (GS) module for Re-ID. SVA is to backtrack and predict the pedestrian trajectory of tail classes, and DVA is to use diffusion model to change the background of the scene. GS divides the pedestrians into unrelated groups and performs softmax operation on each group individually. Our proposed strategies can be integrated into numerous existing tracking systems, and extensive experimentation validates the efficacy of our method in reducing the influence of long-tail distribution on multi-object tracking performance. The code is available at

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper attempts to address the issue of long-tail distribution in pedestrian trajectory data in multi-object tracking (MOT) tasks. Specifically, the authors found that in existing multi-object tracking datasets, there is a significant imbalance in the trajectory lengths of different pedestrians. Some pedestrians have very rich trajectory data (head class), while others have very sparse trajectory data (tail class). This long-tail distribution causes the network to be biased towards learning head class features during training, while ignoring tail class features, thereby affecting the overall performance of multi-object tracking.

To solve this problem, the authors propose the following strategies:
1. **Data Augmentation Strategies**:
- **Static Camera View Data Augmentation (SV A)**: Includes two techniques, backtracking continuation and predictive continuation, to increase the trajectory data of tail class pedestrians.
- **Dynamic Camera View Data Augmentation (DV A)**: Uses diffusion models to change the scene background, enhancing the network's focus on pedestrian features.
2. **Module Improvements**:
- **Group Softmax (GS) Module**: Divides pedestrians into multiple unrelated groups and performs softmax operations on each group separately to prevent the excessive suppression of tail class weights by head class weights, thereby improving the network's ability to extract tail class features.

Through these strategies, the authors aim to reduce the impact of long-tail distribution on multi-object tracking performance and enhance overall tracking effectiveness. Experimental results show that these methods achieve significant performance improvements on multiple public MOT benchmark datasets.