SIU
ACM Digital Library 2024

Structured Instance Understanding with Boundary Box Relationships

We introduce a post-processing validator that reasons over geometric relationships of predicted bounding boxes to detect inconsistent object structures, improving reliability across industrial object detection pipelines.

Supasate Vorathammathorn Thanatwit Angsarawanee Sakol Tasanangam Theerat Sakdejayont
Conference
ACM 2024
Domain
Post-Detection Validation
Keywords
Object Detection, Structural QA
Read the Paper View Code Dataset
Abstract

Structured Consistency for Object Detection

Bounding boxes alone provide limited structural guarantees. SIU elevates detection pipelines by verifying that inferred components respect the spatial grammar of the target system.

We present Structured Instance Understanding (SIU), a validation layer that inspects the topology of bounding boxes emitted by an object detector. The validator records canonical geometric relationships between parts using domain priors, and rejects detection outputs that violate these relationships. Applied to the Car Parts dataset, SIU identifies malformed assemblies with precision between 85–92% and recall between 88–95%, while maintaining minimal runtime overhead. The framework is model-agnostic and can be adapted to other structured detection domains such as PCB inspection or medical imaging workflows.
Highlights

Key Contributions

Relational Validator

Encodes canonical spatial relationships between parts and flags detections that violate structural constraints.

Lightweight Deployment

Adds <1 second runtime overhead, enabling integration in latency-sensitive production systems.

Robust Gains

Delivers up to 25% improvement in structural reliability by reducing misassembled detections 60–80%.

Model Agnostic

Easily plugs into any object detector, allowing teams to elevate quality control without retraining.

Method

Validation Workflow

SIU integrates after the detector, composing a three-phase pipeline that contextualizes bounding boxes before final approval.

1

Graph Construction

Convert bounding boxes into a relational graph where nodes capture object categories and edges encode adjacency, alignment, and distance statistics.

  • Derive canonical anchor points for each object class.
  • Compute directional relationships and scale-normalized gaps.
2

Constraint Evaluation

Apply domain-specific rules to the graph to score structural plausibility using statistical thresholds.

  • Check mandatory part presence and configuration.
  • Measure pairwise deviations against learned tolerances.
3

Decision & Feedback

Flag suspect detections and produce diagnostics to trigger downstream human review or automatic retries.

  • Emit structured failure codes for interpretability.
  • Provide actionable feedback for detector refinement.
Evaluation

Experimental Results

Evaluation on the Car Parts dataset with 3,291 images across 20 categories demonstrates substantial accuracy and reliability gains.

Metric Value Description
Accuracy 85–95% Overall classification accuracy on structural validity labels.
Precision 85–92% Rate of correct structure predictions among accepted detections.
Recall 88–95% True positive rate of structural anomaly detection.
F1 Score 85–92% Harmonic mean capturing overall balance.
ROC AUC 0.90–0.95 Area under the ROC curve across operating points.
60–80%
Error Reduction
15–25%
Reliability Gain
< 1 sec
Computation
Resources

Reproduce & Extend

Get started with the open-source implementation, documentation, and dataset to adapt SIU for your domain.

# Clone the repository git clone https://github.com/ksupasate/SIU-ObjectDetection-Validator cd SIU-ObjectDetection-Validator # Install dependencies pip install -r requirements.txt # Train the SIU model python main.py train # Run inference python main.py inference "path/to/image.jpg"
Explore on GitHub Read Documentation
Reference

Citation

If you use this work in your research, please cite the following. 

@inproceedings{siu2024,
  title     = {Structured Instance Understanding with Boundary Box
               Relationships in Object Detection System},
  author    = {Vorathammathorn, Supasate and Angsarawanee, Thanatwit and
               Tasanangam, Sakol and Sakdejayont, Theerat},
  booktitle = {Proceedings of the ACM Conference},
  year      = {2024},
  doi       = {10.1145/3643487.3662729}
}