data association

Multi-object tracking (MOT) suffers from hand-crafted association rules, limiting simultaneous gains in efficiency and robustness. Method: We propose Track-Detection Link Prediction (TDLP), the first framework to introduce link prediction into the detection-to-tracking paradigm. TDLP performs end-to-end learning of inter-frame associations between trajectories and detections, enabling data-driven joint modeling of geometric and heterogeneous features (e.g., pose and appearance). It employs a lightweight graph neural network backbone that encodes bounding-box geometry and supports optional multimodal feature fusion, optimized via a joint objective combining contrastive learning and structured matching loss. Results: TDLP achieves state-of-the-art performance on MOT17 and MOT20 benchmarks, outperforming both leading detection-based trackers and end-to-end methods. Ablation studies confirm that link prediction significantly surpasses conventional metric learning for heterogeneous feature integration. The code is publicly available.

This paper addresses cross-view identity inconsistency in multi-camera multi-object tracking (MCMT). To resolve this, we propose a trajectory-appearance fusion framework with global-priority identity assignment. Our method integrates BoT-SORT for single-camera tracking, 3D trajectory reconstruction (leveraging depth maps and camera calibration), and cross-view feature matching into an end-to-end global association pipeline. Key contributions include: (1) a Glance initialization mechanism enabling low-latency, robust introduction of new targets; (2) a progressive matching strategy that prioritizes re-identification of existing IDs and creates new IDs only when confidence falls below a threshold, thereby substantially suppressing ID switches; and (3) unified identity assignment via joint spatiotemporal and appearance modeling. Experiments on multiple MCMT benchmarks demonstrate significant improvements in IDF1 and a reduction in ID switches by over 32%, markedly enhancing cross-view tracking consistency and accuracy under complex scenarios.

Multi-View Multi-Object Tracking (MV-MOT) aims to localize and maintain consistent identities of objects observed by multiple sensors. This task is challenging, as viewpoint changes and occlusion disrupt identity consistency across views and time. Recent end-to-end approaches address this by jointly learning 2D Bird's Eye View (BEV) representations and identity associations, achieving high tracking accuracy. However, these methods offer no principled uncertainty accounting and remain tightly coupled to their training configuration, limiting generalization across sensor layouts, modalities, or datasets without retraining. We propose ModTrack, a modular MV-MOT system that matches end-to-end performance while providing cross-modal, sensor-agnostic generalization and traceable uncertainty. ModTrack confines learning methods to just the \textit{Detection and Feature Extraction} stage of the MV-MOT pipeline, performing all fusion, association, and tracking with closed-form analytical methods. Our design reduces each sensor's output to calibrated position-covariance pairs $(\mathbf{z}, R)$; cross-view clustering and precision-weighted fusion then yield unified estimates $(\hat{\mathbf{z}}, \hat{R})$ for identity assignment and temporal tracking. A feedback-coupled, identity-informed Gaussian Mixture Probability Hypothesis Density (GM-PHD) filter with HMM motion modes uses these fused estimates to maintain identities under missed detections and heavy occlusion. ModTrack achieves 95.5 IDF1 and 91.4 MOTA on \textit{WildTrack}, surpassing all prior modular methods by over 21 points and rivaling the state-of-the-art end-to-end methods while providing deployment flexibility they cannot. Specifically, the same tracker core transfers unchanged to \textit{MultiviewX} and \textit{RadarScenes}, with only perception-module replacement required to extend to new domains and sensor modalities.

This paper addresses long-term point tracking in online settings—i.e., establishing cross-frame physically consistent correspondences for target points using only current and past frames (causal setting). We propose Track-On, the first method to integrate vision foundation models with online temporal modeling: it treats each tracked point as a query and processes frames sequentially, dynamically maintaining appearance and contextual memory without accessing future frames or sliding windows. Built upon a Transformer architecture, Track-On jointly performs spatial feature extraction and point-level query-driven progressive temporal reasoning to ensure long-term consistency. Evaluated on seven public benchmarks, Track-On achieves state-of-the-art performance, demonstrating the feasibility of causal long-term point tracking and significantly improving robustness and accuracy in online applications such as robotic navigation and augmented reality.

In city-scale multi-camera vehicle tracking, cross-camera trajectory matching typically relies on labor-intensive spatiotemporal annotations, leading to high deployment costs and poor scalability. Method: This paper proposes a self-supervised camera linking approach that eliminates the need for manual labels. It introduces a joint pre-matching mechanism leveraging only vehicle appearance features, trajectory-pair counts, and temporal variance to dynamically infer high-confidence spatial associations between cameras. Furthermore, it designs a probabilistic camera-linking inference framework for end-to-end camera relationship modeling. Contribution/Results: Evaluated on the CityFlow V2 benchmark, our method achieves an IDF1 score of 61.07%, setting the new state-of-the-art among fully automatic camera-linking methods. It significantly enhances system scalability and practical deployment efficiency while removing dependency on human annotation.

This work addresses the lack of efficient and identity-consistent annotation tools for purely vision-based multi-camera multi-object tracking, where prevailing approaches relying on single-view imagery or LiDAR often fail to ensure cross-view identity coherence. To overcome this limitation, we propose the first unified bird’s-eye-view (BEV)-based annotation framework that enables one-time labeling in the BEV plane and automatically projects these annotations into 2D bounding boxes across all camera views. The system integrates camera calibration, off-the-shelf 2D detectors, and footpoint estimation, and further accelerates the annotation process through temporal propagation and semi-automatic candidate generation. Evaluated on a multi-camera poultry tracking task, our method substantially improves annotation efficiency, significantly reducing manual effort compared to conventional per-camera labeling while rigorously preserving cross-view identity consistency.

Existing video re-rendering methods struggle to simultaneously preserve appearance fidelity, ensure dynamic consistency, and enable precise camera control under novel viewpoints, while lacking explicit spatiotemporal pixel correspondences. To address these limitations, this work proposes a video diffusion Transformer conditioned on paired 3D point trajectories, achieving four-dimensional consistent and camera-controllable generation. The key innovations include a data pipeline that extracts one-to-one trajectory correspondences from multi-view videos and a dual-view trajectory conditioner that integrates geometric operations with temporal aggregation. Evaluated on a benchmark of 400 videos encompassing both static and dynamic scenes, the proposed method substantially outperforms existing approaches, reducing rotation errors by 30–65% and translation errors by 61–72%.

This work addresses the challenges of identity inconsistency and degraded segmentation accuracy in multi-object tracking and segmentation (MOTS), which stem from unreliable trajectory association and error propagation from false detections. To tackle these issues, the paper proposes a zero-shot MOTS approach that integrates the SAM2 instance segmentation model with a novel trajectory management module. The method introduces Mask Centroid Distance (MCD) and Confidence-aware Cost Modulation (CCM) to refine data association and incorporates a Bernoulli filter–based probabilistic trajectory verification mechanism to enhance temporal consistency. Without requiring any fine-tuning, the proposed framework effectively suppresses spurious trajectories and improves identity preservation, achieving significant performance gains in both tracking and segmentation on the KITTI MOTS benchmark.

This work addresses the challenge of misalignment between target reference and trajectory planning in embodied visual tracking under monocular vision and natural language instructions. The authors propose a “reference-then-track” paradigm: first explicitly selecting the linguistically described target in image space, then decoding a tracking trajectory based on this decision. Their method introduces Temporal View Bounding Indicator (TVBI) tokens to fuse geometric features, incorporates a sliding-window memory mechanism to retain historical motion cues, and leverages a newly constructed Refer-QA dataset for joint training of vision–language–action policies. Evaluated on the EVT-Bench single-view setting, the approach achieves state-of-the-art success rates of 89.4%, 73.3%, and 74.1% across three task categories—matching or surpassing multi-camera baselines—and demonstrates strong sim-to-real transferability on both quadrupedal and humanoid robots.

This work addresses weakly supervised cross-view referring multi-object tracking, eliminating the need for per-frame spatial annotations and cross-view identity labels. The authors propose a two-stage framework: first, initial trajectories are generated using SAM2/3, followed by an affinity-guided cross-view re-prompting strategy to produce high-quality pseudo-labels. They then introduce ViewSAM, a model that explicitly captures view-aware cross-modal semantics through a learnable view-conditioning mechanism, effectively aligning view-varying visual observations with view-invariant linguistic expressions. The approach introduces only approximately 10% additional parameters and achieves state-of-the-art performance under weak supervision, remaining competitive even against fully supervised methods.