Source-linked AI summary
Exploring Learning Models for Topological Relationship Recognition from Image Data
Saptak Das, Monidipa Das
TL;DR
Image-based topological relationship classification remains understudied, despite its relevance to fields such as GIS, medical imaging, and computer vision. This study creates an annotated dataset and compares traditional and deep learning models. VGG16 reaches 89.55% validation accuracy, and the authors report that pre-trained CNNs outperform traditional models.
Problem
Image-based topological relationship classification remains a challenging, relatively understudied task, although object relationships matter in fields including GIS, medical imaging, and computer vision.
Source: §1 · p. 1Method
The study builds an approximately 11,000-image dataset labeled for three topological relationships and compares traditional classifiers with neural models.
Source: §1 · p. 2Results
89.55% validation accuracy was reported for VGG16, and pre-trained CNNs outperformed traditional machine-learning models in validation accuracy and generalization.
Source: §4 · p. 9Takeaways & Limitations
The dataset and comparative benchmarks provide a resource for research on automatic topological relationship classification.
Source: §1 · p. 2
Abstract
from arXiv · showhide
Figuring out how objects relate to each other, like whether they touch, overlap, stay completely separate or one sits inside another, matters a lot in fields like GIS, biomedical imaging, and robotics. Even though machine learning has come a long way, people haven't really focused on spotting these topological relationships in images. The main roadblocks? Not enough good datasets and no clear way to measure results. So, we rolled up our sleeves and built a new dataset. It's pretty sizable: over 11,000 labelled images showing all those essential relationships. We ran tests with some classic machine learning models, Naive Bayes, KNN, Random Forest, SVM, and Artificial Neural Networks, and threw in some deep learning stars like VGG16 and InceptionResNetV2. For the dataset itself, we used segmentation, contour detection, and grayscale normalization to tease out solid feature vectors. The results? Deep learning methods, especially VGG16, pulled ahead, with validation accuracy hitting 89.55%. That's a big jump compared to the traditional models. This shows how powerful transfer learning is for analyzing topological relationships in images, and it gives researchers a new standard to aim for in future work on spatial reasoning and topological classification.
1 Introduction
Topological relationships matter across spatial applications, but their image-based classification remains challenging and lacks large-scale datasets and benchmarks. This study introduces an annotated dataset and compares traditional and deep learning classifiers to establish performance benchmarks.
1 Introduction
- The study targets a literature gap: image-based topological relationship classification has received little attention, with no large-scale datasets or benchmark models reported.The authors connect the task to GIS, medical imaging, computer vision, and applications including robotics and biomedical analysis.Source: §1 · pp. 1–2
- The proposed dataset contains around 11,000 images annotated for Touch/Overlap, Disjoint, and Containment.The dataset is intended as a resource for training and evaluating models, and is to be made public upon acceptance.Source: §1 · p. 2
- The study compares classical classifiers and neural models, including CNNs, ANNs, Naïve Bayes, KNN, Random Forest, and SVM, to set benchmarks.The authors frame the comparison as a contribution to automatic topological relationship recognition.Source: §1 · p. 2
- The authors connect more accurate relationship classification to spatial-data interpretation, grasping and manipulation, and biomedical image analysis.They identify GIS and cartography, robotics, and segmentation of cells, tumours, or anatomical structures as relevant settings.Source: §1 · p. 2
2 Related Works
Prior work includes rule-based and hand-crafted approaches, a widely used nine-intersection framework, and some CNN-based spatial reasoning. The paper identifies limited formal topological labeling in available datasets and contributes a dataset and model benchmarks for this classification task.
2 Related Works
- Rule-based heuristics and manually created features offered insight but had scaling and reliability issues in complex real-world settings.The paper situates these approaches among earlier methods.Source: §2 · p. 3
- The nine-intersection model by Egenhofer and Herring describes basic relations such as Disjoint, Meet, Overlap, and Contain and is used in spatial reasoning and query optimization.The passage also notes its use in indexing approaches for spatiotemporal data.Source: §2 · p. 3
- CNNs have reported strong results in vision tasks, but their use for topological relationship classification remains at an early stage.Prior spatial-reasoning work has focused more on geometric proximity and functional relationships than pure topological semantics.Source: §2 · p. 3
- The paper identifies a lack of datasets that properly label topological relationships; Visual Genome emphasizes spatial interactions but does not fully cover formal topological structures.The authors present a new annotated dataset and model evaluation as a response to these gaps.Source: §2 · p. 3
- The paper introduces a dataset labeled for Touch/Overlap, Disjoint, and Containment and evaluates machine-learning and deep-learning models.It presents this work as encouraging further development in automatic topological relationship classification.Source: §2 · p. 3
3 Methodology
The methodology prepares images and assigns topological labels from object contours, then trains and evaluates classical and deep-learning classifiers on the resulting feature data.
3.1 Image Processing and Image Labeling:
- Images are resized, segmented by K-means with K = 3, grayscale-converted, thresholded, and morphologically closed before contour-based analysis.The authors say these preprocessing steps improve contour detection and the reliability of spatial-relationship analysis.Source: §3.1 · p. 3
- Contour-based labels distinguish Containment, Touch/Overlap, and Disjoint according to whether contour points satisfy the defined neighborhood relationship.Containment requires all contour points of one object to be neighbors of the other’s; partial satisfaction yields Touch/Overlap, and none yields Disjoint.Source: §3.1 · p. 5
- The workflow proceeds from preprocessing and labeling through feature-vector generation and dataset partitioning to model training, tuning, and evaluation.Pixel-level grayscale values from resized 24×24 images form the structured feature data.Source: §3.1 · p. 5
3.2 Machine Learning and Deep Learning Models:
- The dataset was divided into training (60%), validation (20%), and testing (20%) subsets; feature values were normalized to [0, 1] by dividing by 255.0, improving neural-network training accuracy and convergence speed.Source: Dataset Description · p. 5
- Neural models use Adam and categorical crossentropy for 100 epochs; neural-network tuning uses Keras Tuner random search, while classical models use stratified five-fold RandomizedSearchCV.VGG16 uses frozen base layers and augmentation; InceptionResNetV2 uses early stopping without tuning its base layers, and Naïve Bayes is untuned.Source: Dataset Description · pp. 5–6
- Models are evaluated separately on training, validation, and test sets using accuracy, confusion matrices, and classification reports with precision, recall, and F1-scores.The confusion matrices are used to examine class-wise prediction accuracy and misclassifications.Source: Model Evaluation · p. 7, Figure 2 · p. 8
4 Experimental Study
The study compares classical and deep-learning classifiers on the dataset, reporting stronger validation accuracy and generalization for pretrained CNNs. VGG16 leads the reported validation results, while the authors identify future model and domain extensions.
4 Experimental Study
- 89.55% validation accuracy for VGG16; Random Forest and SVM validation accuracies were in the 75% range.The results table reports model performance; the prose also notes high CNN accuracy across splits.Source: Table 3 · p. 9, §4 · p. 9
- KNN reaches perfect training results but overfits, while Naïve Bayes and ANN without dropout perform weakly and are described as unable to model dataset complexity.The authors also emphasize the importance of model selection and hyperparameter tuning.Source: §4 · p. 9, §5 · p. 10
- Random-search cross-validation improves tuned classical-model performance; all classical models except Naïve Bayes were tuned, which the authors describe as ensuring a fair comparison.Random Forest and SVM are specifically reported to have validation accuracies in the 75% range.Source: §4 · p. 9
- Pretrained VGG16 and InceptionResNetV2 outperform traditional machine-learning models in validation accuracy and generalization.The study compares training, testing, and validation performance across the models.Source: §4 · pp. 8–9
- The authors frame the study as a foundation for future algorithm comparisons and propose exploring more complex architectures, multimodal data, and domain-specific improvements.They identify GIS, biomedical imaging, and robotics as application areas dependent on accurate spatial-relationship analysis, and mention future MRI and CT work.Source: §5 · p. 10