Representation Alignment Contrastive Regularization for Multi-Object Tracking

TL;DR

This work tackles multi-object tracking data association by introducing a detector-free, lightweight Representation Alignment Module (RAM) built on a single-layer transformer to encode spatio-temporal relationships. RAM is guided by two interpretability-driven contrastive regularizations that formalize spatial and temporal alignment as triplet-based losses, with alignment fused into the data association via $A_T = \alpha_T S(\mathcal{H}^t, \mathcal{C}^{t-1}) + (1-\alpha_T) S(\bar{\mathcal{H}}^t, \bar{\mathcal{C}}^t)$ and training losses $L_T$, $L_S$, $L_{ST}$. The approach yields three RAM variants—TRAM, SRAM, and STRAM—and demonstrates consistent improvements across MOT17, MOT20, and BDD100K when integrated with popular trackers, with minimal training overhead and even enabling unsupervised use. The results underline RAM’s practical impact: improved association accuracy and reduced identity switches with little cost, making it a scalable enhancement for diverse MOT systems; code is provided for reproducibility.

Abstract

Achieving high-performance in multi-object tracking algorithms heavily relies on modeling spatio-temporal relationships during the data association stage. Mainstream approaches encompass rule-based and deep learning-based methods for spatio-temporal relationship modeling. While the former relies on physical motion laws, offering wider applicability but yielding suboptimal results for complex object movements, the latter, though achieving high-performance, lacks interpretability and involves complex module designs. This work aims to simplify deep learning-based spatio-temporal relationship models and introduce interpretability into features for data association. Specifically, a lightweight single-layer transformer encoder is utilized to model spatio-temporal relationships. To make features more interpretative, two contrastive regularization losses based on representation alignment are proposed, derived from spatio-temporal consistency rules. By applying weighted summation to affinity matrices, the aligned features can seamlessly integrate into the data association stage of the original tracking workflow. Experimental results showcase that our model enhances the majority of existing tracking networks' performance without excessive complexity, with minimal increase in training overhead and nearly negligible computational and storage costs.

Figures (8)

• Figure 1: Demonstration of representation alignment rules: Temporal rule reduces gap between consecutive frame's target representations, while spatial rule unites representations of the same object.
• Figure 2: The general process of the RATracker and diagram of contrastive regularization derived from representation alignment rules. (a) Diagram of contrastive regularization terms guided by alignment rules, operator + means weighted sum. (b) Structure of RAMs, operator ⓢ means sequence stack, letters with overbar represent aligned features.
• Figure 3: The average performance of RAMs on various trackers in Table \ref{['tab:flexibility_validation']}.
• Figure 4: The impact of hyperparameter $\lambda$ on STRAM's performance.
• Figure 5: Comparison of rule-based, deep learning, and our methods on MOT17 validation set. Notable tracking errors emphasized.