Quality Matters: Embracing Quality Clues for Robust 3D Multi-Object Tracking

1 Introduction

• Camera-only 3D MOT locates objects and associates detections of the same identities into tracklets, using cues such as position, size, velocity, and appearance.The paper focuses on camera-only tracking because it contains semantic information and is more economical.Source: §1 · p. 1
• The CV tracking framework predicts location from detected centers and velocities, then uses L2 distance for association, but depends heavily on prediction quality.The paper describes this approach as computationally simple, using matrix addition and multiplication for parallel cost computation.Source: §1 · p. 2
• Occlusion, motion blur, and external illumination disturb estimates; location and velocity errors differ substantially and their predicted qualities are misaligned.The authors argue that these imbalanced cues matter greatly for MOT, although they have little effect on detection performance.Source: §1 · p. 2
• NGQ measures predicted location and velocity quality, which QOA uses to filter low-quality motion candidates and reject poorly estimated box centers.QTrack combines these components with a baseline 3D detector.Source: §1 · p. 2
• The authors analyze location and velocity quality distributions and their relationship to explain why efficient CV tracking can remain suboptimal.They propose quality prediction as a way to give each box self-diagnosis for tracking association.Source: §1 · p. 2
• +2.2% AMOTA: QOA improves the baseline tracker across several 3D detector settings, and QTrack outperforms other camera-based methods on nuScenes by a large margin.The paper reports these results as evidence for the quality-aware association strategy and overall tracker.Source: §1 · p. 2

2.1 3D Multi-object Tracking

• Tracking-by-detection methods detect 3D objects and associate them with trajectories using motion, appearance, temporal, or transformer-based cues, but complex visual conditions can degrade tracking.The paper argues that a simple velocity clue paired with quality estimation can address difficult cases.Source: §2.1 · p. 2

2.2 Prediction Quality Estimation

• Prior quality-estimation work predicts box centerness or IoU, mask quality, or depth-factor uncertainty; this paper estimates location and velocity quality for 3D MOT association.Figure 2 describes irregular location and velocity quality distributions and a messy scatter plot reflecting no relation between them.Source: §2.2 · pp. 2–3, Figure 2 · p. 3

2.3 Multi-View 3D Object Detection

• Multi-view camera detectors use 3D object queries or project image features into BEV, and this work equips BEVDepth to assess velocity and center-location quality for tracking.The resulting QTrack is presented as a tracking-by-detection framework for multi-view camera scenes.Source: §2.3 · p. 3

3.1 Delve into the Quality Distribution

• Scattered location and velocity errors, including nonlinear quality misalignment, motivate considering both qualities when selecting association candidates.The analysis identifies low quality in either prediction and misalignment between the two as reasons velocity-based tracking performs poorly.Source: §3.1 · p. 3

3.2 Base 3D Object Detector

• BEVDepth provides the camera-based detector backbone, with regression outputs including heatmap, offsets, height, size, rotation, and velocity.The method can be coupled with existing 3D detectors; BEVDepth is the example used here.Source: §3.2 · pp. 3–4

3.3 Location and Velocity Quality Learning

• Normalized Gaussian Quality (NGQ) represents offset and velocity quality, with predicted-versus-ground-truth error mapped to a score closer to zero as error increases.Separate branches predict location and velocity quality, and the end-to-end training objective adds quality loss to detector loss.Source: §3.3 · p. 4

3.4 Quality-aware Object Association

• QOA first associates detections using classification scores, then rechecks matches between unmatched tracks and low-score detections using location and velocity quality.The association uses velocity-based location prediction, L2 distance, and the Hungarian algorithm.Source: §3.4 · p. 4

4.1 Datasets and Metrics

• nuScenes provides 1,000 real-world sequences, with 700 for training, 150 for validation, and 150 for test; tracking evaluation includes AMOTA and AMOTP.The benchmark uses six cameras with 360-degree coverage and 2 FPS keyframes.Source: §4.1 · p. 5

4.2 Implementation Details

• QTrack uses ResNet-50, ResNet-101, or VoVNet-99 image backbones, with AdamW at a 2 × 10−4 learning rate and batch size 64.The training schedule is 24 epochs for ResNet and 20 for VoVNet with CBGS.Source: §4.2 · p. 5

4.3 Comparision with Preceding SOTAs

• AMOTA reaches 0.511, and QTrack outperforms camera-based state-of-the-art methods on nuScenes validation and test sets.The comparison uses a VoVNet-99 backbone initialized from DD3D and 640 × 1600 input images.Source: §4.3 · p. 5, Table 1 · p. 6

4.4 Ablation Study

• Both QOA gates improve tracking across the tested backbone and frame settings, while classification-confidence gating alone increases IDS.The quality-score gate provides a finer-grained association reference after confidence-based candidate filtering.Source: §4.4 · p. 6, Table 3 · p. 7
• Using location quality or velocity quality alone does not improve tracking, consistent with the finding that their prediction qualities are not aligned.The study therefore treats both scores as reference cues in QOA.Source: §4.4 · p. 6, Table 4 · p. 7
• Adding the quality branch does not affect base 3D detection or CV and SimpleTrack tracking, whereas an appearance branch slightly degrades detection and reduces tracking performance by nearly 1.0%.The reported detection degradation is nearly 0.5%.Source: §4.4 · pp. 6–7, Table 5 · p. 7

4.5 Discussion and Future Work

• In NMS, the threshold metric M combines velocity quality V and classification score C, with α controlling V’s contribution.As V’s contribution increases, mAVE drops while mAP and mATE degrade; NDS first improves and then worsens, reflecting a location-error/velocity-error trade-off.Source: §4.5 · p. 7, Figure 4 · p. 7
• The discussion identifies simultaneous improvement of location or 3D-box and velocity prediction as a challenge for detection and downstream tasks such as 3D MOT.This challenge is linked to the reported misalignment between location and velocity prediction quality.Source: §4.5 · p. 7

• The conclusion identifies imbalanced location and velocity prediction quality as a focus for 3D detection and tracking research.Source: §5 · p. 7
• NGQ measures predicted location and velocity quality, while an effective module learns those qualities.Source: §5 · p. 7
• QOA is proposed to alleviate the location–velocity quality imbalance in 3D MOT.Source: §5 · p. 7
• QTrack incorporates QOA in a tracking-by-detection framework for multi-view camera 3D MOT.Source: §5 · p. 7
• The authors report that extensive experiments demonstrate the method’s efficacy and robustness.Source: §5 · p. 7