[Paper Review] Multi-Target, Multi-Camera Tracking by Hierarchical Clustering: Recent Progress on DukeMTMC Project

Zhimeng Zhang, Jianan Wu+2 more|arXiv (Cornell University)|Dec 27, 2017

Video Surveillance and Tracking Methods5 references65 citations

TL;DR

The paper shows that simple hierarchical clustering guided by high-quality person re-identification features can achieve strong multi-target, multi-camera tracking on DukeMTMC, combining detection, re-id features, single-camera tracking, and hierarchical clustering without distance matrix updates.

ABSTRACT

Although many methods perform well in single camera tracking, multi-camera tracking remains a challenging problem with less attention. DukeMTMC is a large-scale, well-annotated multi-camera tracking benchmark which makes great progress in this field. This report is dedicated to briefly introduce our method on DukeMTMC and show that simple hierarchical clustering with well-trained person re-identification features can get good results on this dataset.

Motivation & Objective

• Motivate and evaluate multi-target, multi-camera tracking (MTMC) on the DukeMTMC benchmark.
• Propose a simple, scalable MTMC approach based on hierarchical clustering of tracklets and trajectories.
• Demonstrate that high-quality re-identification features can drive strong MTMC performance.
• Assess the impact of private vs public detections and re-ranking on MTMC performance.

Proposed method

• Detect people with Faster R-CNN and a Confidence/IoU thresholding pipeline.
• Extract appearance features with a person re-id model (AlignedReID style) trained on multiple public datasets.
• Perform near-online single-camera tracking by merging adjacent frame detections into tracklets via Kuhn–Munkres data association, then hierarchically cluster tracklets into trajectories within each camera.
• Merge trajectories across cameras using hierarchical clustering without distance matrix updating, after averaging re-id features across trajectories and applying re-ranking to the distance matrix.
• Apply simple temporal and cross-camera constraints (e.g., max one minute separation, limited simultaneous appearances) to refine cross-camera merges.

Experimental results

Research questions

• RQ1 Can hierarchical clustering with robust re-id features achieve competitive MTMC performance on DukeMTMC?
• RQ2 How do private vs public detections and re-ranking influence MTMC performance?
• RQ3 What is the effect of single-camera tracking quality on downstream multi-camera linkage?
• RQ4 What practical constraints improve cross-camera trajectory merging without compromising scalability?

Key findings

• On easy test set with public detections, MTMC_ReIDp achieves IDF1 74.4, IDP 84.4, IDR 66.4.
• On easy test set with private detections, MTMC_ReID achieves IDF1 83.2, IDP 85.2, IDR 81.2.
• On hard test set with public detections, MTMC_ReID achieves IDF1 74.0, IDP 81.4, IDR 67.8 (comparison with DeepCC 68.5/75.9/62.4).
• On hard test set with private detections, MTMC_ReID achieves IDF1 74.0, IDP 81.4, IDR 67.8 (DeepCC 68.5/75.9/62.4).
• Re-ranking of the re-id distance matrix provides about a 2–3 percentage-point gain in IDF1 on training/validation.
• Overall, the simple hierarchical clustering framework with strong re-id features achieves state-of-the-art results across configurations.