Source-linked AI summary
RailSyn: Diagnosis-Guided Image Generation for Traceable Data Completion in Railway Foreign Object Detection
Quan Hao, Chenxi Zhang, Ziyang Tao, Yuyuan Zhou, Yudong Wang, Rui Shi, Lechuan Xu, Changhao Liu, Liguo Zhang
TL;DR
Scarce real RFOD positives incompletely capture task-relevant railway context, intrusion relations, and visual variation. RailSyn diagnoses these gaps and generates traceable completion data, improving AP50–95 across nine detectors by up to 4.9 points while reaching 13.64% local-gap occupation.
Problem
Scarce RFOD positives incompletely represent railway context, intrusion semantics, object scale, and environmental variation needed for reliable detection.
Method
RailSyn combines a real-referenced Inspector for deficiency analysis with a requirement-aligned Generator using domain adaptation, relation-aware planning, and conditional refinement.
Results
AP50–95 improves across all nine tested detectors, with a maximum gain of 4.9 points and full-configuration local-gap occupation Cgap = 13.64%.
Takeaways & Limitations
RailSyn supplies detection-relevant railway context, intrusion relations, and object–background cues that remain useful across heterogeneous detector architectures.
Takeaways & Limitations
Cross-dataset effectiveness remains difficult to characterize, and generation quality requires further improvement, especially under adverse weather and for small-object details.
Abstract
from arXiv · showhide
Railway foreign object detection (RFOD) is critical to safe railway operation, yet scarce real positive samples incompletely represent task-relevant variations in object scale, intrusion relation, railway scene, illumination, and adverse weather. Existing synthetic augmentation can improve RFOD detection, but its gains lack an explicit account of the task-relevant deficiencies complemented by the generated data. We therefore introduce RailSyn, a diagnosis-guided framework comprising a real-referenced Inspector and a requirement-aligned Generator. The Inspector constructs a variable-radius empirical cover from finite real observations to localize candidate completion regions and profile synthetic pools. The resulting audit identifies railway-context, intrusion-semantic, and visual-consistency requirements; the Generator addresses them through domain adaptation, agent-planned placement and physical contact relations, and plan-consistent conditional refinement. Using the Inspector, we further trace representation-space changes across generation variants; the complete system attains a local-shell occupation of $C_{gap}$ to 13.64%, which measures generated coverage of real-derived completion regions. Extensive experiments show AP50--95 gains of up to 4.9 points and consistent improvements across nine mainstream detectors, demonstrating broad cross-architecture utility.
Introduction
RailSyn introduces an Inspector–Generator framework that diagnoses real-referenced deficiencies in RFOD data and guides traceable synthetic completion. Its complete system improves detection across nine mainstream detectors while attaining Cgap = 13.64%.
- Motivation: Finite RFOD datasets incompletely capture railway context, intrusion semantics, visual appearance, and local object–background effects relevant to detection.Reliable detection requires jointly representing railway structures, environmental conditions, object scales, and physical intrusion relations.
- Framework: RailSyn connects real-referenced deficiency analysis, guided synthesis, and representation-space assessment for traceable RFOD data completion.The Inspector evaluates synthetic pools using reliable-support coverage (Crel), local-gap completion (Cgap), and nonredundant volume efficiency (ηvol).
- Inspector: The Inspector identifies high-uncertainty regions from finite real references and profiles nearby synthetic samples using a real-derived local-shell criterion for Cgap.Figure 1 describes recurring gaps in railway-background fidelity, object–scene intrusion semantics, and anomalous injected-object edge effects.
- Generator: The Generator addresses diagnosed gaps through railway-domain adaptation, Agent-planned placement and physical-contact relations, and plan-consistent conditional refinement.These processes target structural backgrounds, background-conditioned executable intrusion plans, and anomalous boundary effects.
- Results: 3.03 points is the average best observed AP50–95 gain across nine mainstream detectors, while 4.9 points is the maximum gain.The complete Generator attains Cgap = 13.64%, and Inspector diagnostics show trends broadly consistent with AP changes in the ablation study.
Related Work
Related work frames RFOD as a safety-critical, small-sample detection problem addressed by mainstream YOLO- and DETR-style detectors and data-centric augmentation. Recent generative methods provide high-resolution synthesis, spatial control, and improved structure preservation for limited-observation settings.
- RFOD detects debris and intrusions near railway tracks, where missed objects can obstruct operations and threaten safety.
- YOLO-style detectors prioritize efficient multi-scale detection, whereas DETR-style models use transformer-based matching to exploit global context.
- Rare intrusions and costly acquisition leave RFOD with few positive samples, weakening learning across object scale, illumination, weather, and intrusion relations.
- Data-centric augmentation and generation offer a practical route for supplementing scarce RFOD training samples.
- Latent diffusion supports high-resolution synthesis in compressed spaces and augmentation under limited observations, while flow matching and FLUX use rectified-flow Transformers.
- Spatial controls improve generated-object placement and structure preservation.
Method
RailSyn treats synthetic RFOD augmentation as traceable completion of finite real observations. Its Inspector identifies local completion regions and deficiencies, while the requirement-aligned Generator addresses railway context, intrusion relations, and visual consistency.
- Inspector: The Inspector constructs real-referenced completion regions, profiles synthetic pools, and separately evaluates completion quality and real-scene detection utility.It assesses both what generated data contribute beyond finite real observations and whether detectors exploit that contribution.
- Inspector: The Inspector quantifies reliable-support reach, local-shell occupation, and generated-cover nonredundancy using overlap-corrected native-spherical Monte Carlo.High-mass shells are ranked for real-reference and RFOD23 AIGC candidate retrieval, enabling audits independent of downstream generation choices.
- Inspector: Inspector audits identify railway-context, intrusion-semantic, and intrusion-pattern deficiencies in generated RFOD samples.Observed issues include disordered non-authentic railway backgrounds and implausible foreign-object positions relative to track infrastructure.
- Generator: The Generator addresses diagnosed deficiencies through railway-scene and object LoRAs, relation-aware intrusion planning, and diagnosed intrusion-pattern refinement.The coupled process preserves one-to-one correspondence between inspected requirements and generated realizations.
- Generator: A multimodal Agent couples object category, placement, physical relation, contact region, and tilt, propagating accepted plans to annotations and conditional realization.Plans explicitly reference rails, sleepers, ballast, or catenary and are normalized to valid image coordinates and scale ranges.
- Generator: Multiscale control-map fusion transfers planned intrusion patterns from global extent to local contact and contours while suppressing anomalous boundaries.Zero initialization preserves the pretrained generation path initially, and softened alpha composition maintains object visibility.
Experiments
Experiments show that RailSyn improves detection across architectures and augmentation scales, with complete-stage generation outperforming alternatives. Inspector diagnostics trace task-relevant coverage, correlate with detector utility, remain stable across encoders, and extend to adverse-weather completion.
- Cross-architecture scaling: Every one of nine detectors exceeds its real-only AP50–95 baseline at a suitable RailSyn augmentation scale, with gains averaging 3.03 points and reaching +4.9 points.YOLO11 rises from 39.6 to 42.4 AP50–95 at +300 synthetic samples; optimal augmentation scales vary by architecture.
- Stage ablation: The complete three-stage configuration improves all four primary AP entries, raising AP50/AP50–95 by 4.7/2.0 points for YOLO11 and 5.1/2.9 points for DEIM.Partial configurations benefit one detector while harming or inconsistently affecting the other, demonstrating the need for scene preparation, relation planning, and pattern refinement together.
- Inspector traceability: Cgap increases from 11.88% to 12.82% and ηvol from 75.90% to 77.13% when intrusion-pattern refinement is added, while Crel decreases from 2.77% to 2.73%.Variants with higher Cgap correspond to the best overall AP, whereas FID and KID favor incomplete configurations and miss stage- and scale-specific utility shifts.
- Augmentation-pool comparison: RailSyn achieves 77.1/49.1 on DEIM and 67.3/41.6 on YOLO11, ranking best in all four AP columns with a largest mean improvement of +3.7 points over real-only training.External pools include SD, Nano500, RFOD23, SODA500, and STL500; their gains are smaller or negative.
- Diagnostic validation: Crel and Cgap correlate positively with all four AP measures, with strongest associations of ρ = .735 for DEIM AP50 and ρ = .667 for YOLO11 AP50–95.Across Qwen, CLIP, and DINOv2, Cgap correlations range from 0.771 to 0.943, while ηvol ranges from 0.943 to 1.000.
- Adverse-weather extension: RailSynWeather raises DEIM AP50 from 72.6% to 73.1% on the clean validation split while generating 200 weather-specific images and preserving layout and annotations.The extension uses the same 398-image real training set and addresses underrepresented weather conditions.
Conclusion
RailSyn combines an Inspector–Generator framework with scene preparation, relation-aware planning, and intrusion-pattern refinement to complete high-uncertainty regions in real RFOD observations. Its full configuration achieves the highest local-gap occupation, while cross-dataset characterization and generation fidelity remain open challenges.
- Contribution: 13.64% local-gap occupation is achieved by RailSyn’s full configuration, which also enables component-level analysis of generation stages and scale choices.The framework diagnoses high-uncertainty regions from real RFOD observations and traces how each generation stage and scale choice changes completion profiles.
- Contribution: RailSyn generates synthetic data through scene preparation, relation-aware planning, and intrusion-pattern refinement within its Inspector–Generator design.The Inspector diagnoses high-uncertainty regions from real RFOD observations before generation.
- Limitations and Future Work: Future work targets cross-dataset evaluation, extension to other safety-critical domains, and improved fidelity under adverse weather and for small-object details.Current limitations are the difficulty of characterizing cross-dataset effectiveness and the need for further improvement in generation quality; proposed evaluation uses token-level semantic probes.