P-SRM: Selective Recovery of Rejected Predictions in Visual Tracking

1. INTRODUCTION

• TrackNet suppresses weak heatmap responses, TAP-Net and TAPIR use visibility or localization uncertainty, and this KCF implementation thresholds correlation-filter response peaks.These mechanisms separate accepted outputs from rejected candidates in their respective tracking tasks.Source: §1 · p. 1
• Native rejection can withhold a correctly localized candidate when its confidence or heatmap peak falls below a threshold, while rejected candidates also include mislocalizations and predictions made when the target is invisible.The challenge is to distinguish correctly localized candidates from actual errors.Source: §1 · p. 1
• Prior work calibrates confidence, scores masks or flow correspondences, uses low-score detections for association, or jointly learns rejection; this paper instead verifies candidates after native rejection.The stated question is whether the original tracker’s evidence can identify and recover correctly localized candidates while preserving accepted outputs and candidate coordinates.Source: §1 · p. 1
• P-SRM combines adapted spatial evidence, motion history, and native decision margins to reassess rejected candidates and selectively readmit reliable predictions.The method is introduced as post-rejection candidate verification.Source: §1 · p. 1
• P-SRM identifies and recovers correctly localized rejected candidates while preserving native accepted outputs and candidate coordinates, with identification and tracking benefits validated across six trackers and four datasets.The experiments span category-specific, point, and generic object tracking.Source: §1 · p. 1

2.1. Problem setting and recoverable candidates

• With candidate coordinates fixed, P-SRM treats visible, correctly localized candidates as recoverable and distinguishes them from mislocalized or invisible candidates.For point candidates, correctness uses a localization tolerance; box candidates require IoU of at least 0.5.Source: §2.1 · p. 2

2.2. Spatial evidence adaptation and quality learning

• The adapter aligns candidate locations with same-pass tracker evidence to produce response, candidate-support, and valid-support planes, then combines them with candidate geometry for admission scoring.Geometry includes normalized candidate coordinates and indicators for in-frame location and support boundaries.Source: §2.2 · p. 2
• Training supervises candidate correctness and target presence, then refines ranking with a KL-DRO partial-AUC objective that emphasizes larger pairwise errors.Only the admission logit enters fusion.Source: §2.2 · p. 2

2.3. History and native decision fusion

• P-SRM combines spatial quality with descriptors of motion consistency from past accepted states and the candidate’s native margin relative to the admission boundary.History features use past accepted positions and confidence information; unavailable terms are zero-filled, with flags for the number of past acceptances.Source: §2.3 · pp. 2–3

2.4. Selective recovery

• With the quality network fixed, sequential readouts are fitted using video-grouped five-fold cross-fitting, then refitted on source fitting data for external evaluation.A threshold selected from pooled out-of-fold scores is fixed for readmission; native acceptances are preserved and admitted candidates retain their original coordinates.Source: §2.3 · p. 3, §2.4 · p. 3

3.1. Benchmarks and evaluation protocol

• The evaluation covers category-specific, point, and generic object tracking, with source-data fitting and calibration for TrackNet and point trackers and video-grouped five-fold evaluation for KCF.APr measures identification of correct candidates within the native rejection set; final tracking metrics measure benefits after selective admission.Source: §3.1 · p. 3

3.2. Rejected-candidate identification and recovery

• APr improves across all nine configurations, and pooled F1 gains over Native have 95% paired-bootstrap confidence intervals entirely above zero.Every native rejection set contains correctly localized candidates, which P-SRM recovers without changing their coordinates.Source: §3.2 · p. 4

3.3. Ablation and sensitivity analysis

• Full outperforms M and M+H in ranking and recovery for both point trackers; simpler cue combinations can nevertheless achieve better final tracking metrics in some settings.Comparisons with Q, Q+H, and Q+M also support complementary cue value for ranking; Full is adopted for its applicability across settings and demonstrated gains.Source: §3.3 · p. 4
• Tracking performance remains broadly stable across tested hyperparameter settings, with consistent gains over Native.Sensitivity is assessed in TAP-Net/Kinetics and TrackNetV3/Shuttlecock.Source: §3.3 · p. 4

• Across multiple trackers and datasets, P-SRM improves identification and tracking by recovering correctly localized predictions after native rejection.The paper presents post-rejection verification as a practical way to reuse existing predictions.Source: §4 · p. 4

• The study uses public datasets and existing annotations, adding supplementary target-visibility labels for OTB2013 without new recruitment or video recording.The Shuttlecock Test uses corrected labels released with TrackNetV3.Source: COMPLIANCE WITH ETHICAL STANDARDS · p. 5