Unsupervised Clustering of Public Transport Vehicles Using Deep Morphological Features and Color Histograms

arXiv:2610.22270 · cs.CV · Submitted 2026-10-15 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Unsupervised Clustering of Public Transport Vehicles Using Deep Morphological Features and Color Histograms".

Jane: The gist:

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Moving on to the title and authors of this paper, "Unsupervised Clustering of Public Transport Vehicles Using Deep Morphological Features and Color Histograms." The name itself tells you exactly what they're trying to do: use unsupervised clustering on public transport vehicles.

Jane: It sounds very specific, almost like they are solving a real-world problem right away. The authors are tackling the fact that buses from different routes often look structurally similar but have different paint schemes.

Lu: They’re taking this idea—that morphology and color matter—and combining it with a frozen DINOv2 Vision Transformer for the shape data and handcrafted HSV color histograms for the color data.

Meng: So they are using a pre-trained model just to extract features, not retraining it, which makes sense if you want something fast to run. What’s the actual cost of running that DINOv2 extraction on a live video feed?

Lalam: The paper says this method was evaluated on seven thousand nine hundred seventy-two bus images extracted from about one thousand four hundred sixty hours of video of an arterial road in Mexico City <ref:2610.22270#pg1>. It's testing on a large scale dataset without any labels.

Tom: So the main thing here is that they are aiming to isolate the buses of the city’s public bus network, which they call the RTP, without needing supervision.

Jane: That's huge because it means we can get automated monitoring of transit routes just by looking at their visual properties without needing people to manually label every single vehicle.

Lu: The paper emphasizes that this entire process is unsupervised; no labels are required at any stage of feature extraction or clustering <ref:2610.22270#pg1>.

Meng: That’s the main selling point for deployment because it bypasses the massive effort of manual annotation that would be needed if we tried to build a supervised classification model from scratch.

Lalam: They also used L2-normalized versions of both descriptors before fusing them, which keeps the scales balanced when they combine them <ref:2610.22270#pg1>.

Tom: Exactly. So it’s not just about getting two features; it’s about how you combine those features and then how you use those combined vectors to find meaningful clusters.

Jane: It's a complete pipeline from image input all the way to clustering, which is what makes this approach so powerful for this kind of problem.

Lu: This framework shows that representation learning and traditional computer vision techniques can be effectively integrated when you design the fusion mechanism carefully.

Meng: So if we can get a good starting point without labels, it drastically reduces the initial workload for building a system to identify public transport routes.

The paper's summary: Tom: Now let's talk about what the authors suggest as improvements in "Unsupervised Clustering of Public Transport Vehicles Using Deep Morphological Features and Color Histograms." They point out that they explored different fusion weights alpha to see how they affect the clustering performance.

Jane: So, one key improvement is showing that a moderate color contribution increases the precision of the target cluster with respect to morphological features alone.

Lu: That’s because it proves you can get better performance by combining structural and chromatic information, but only if you tune that weight correctly <ref:2610.22270#pg2>.

Meng: They also showed what happens when alpha gets too high, which leads to oversegmentation or the collapse of the clustering space, so they identified the limits of how much color we can rely on.

Lalam: The paper suggests that color should complement morphology rather than replace it because over-prioritizing color degrades the structural understanding of the vehicles <ref:2610.22270#pg2>.

Tom: That’s a good point because it prevents you from just throwing in a lot of color features and hoping for the best without considering what breaks at scale.

Jane: It gives us a clear guideline on how to balance those two modalities so we don't end up with models that are too brittle when deployed in the real world.

Lu: The paper’s suggestion is that you need to maintain robustness against feature collapse by ensuring color complements the morphological representation rather than replacing it entirely <ref:2610.22270#pg2>.

Meng: From a deployment view, that means we can build systems that are less likely to break if they encounter vehicles with unusual lighting or extreme paint colors.

Tom: It suggests a clear path for tuning this system; start with a baseline and then iteratively adjust alpha to maximize the accuracy you need for your specific task.

Jane: So, in short, the improvement is that it gives us a tunable mechanism to balance structure and color information effectively for unsupervised vehicle clustering.

Lu: It moves the field forward by proving that this hybrid approach is a viable path for solving complex visual grouping problems without needing labeled data.

The paper's improvements: Tom: So, wrapping up on "Unsupervised Clustering of Public Transport Vehicles Using Deep Morphological Features and Color Histograms," we've seen how combining frozen DINOv2 features with HSV color histograms through weighted early fusion works with UMAP projection and K-Means clustering.

Jane: The main implication is that systems can now automatically group public transport vehicles into categories like transit routes without needing manual labels, which is a significant step toward automated monitoring.

Lu: It establishes a solid methodology for representation learning to handle complex visual data where different classes look similar but need to be distinguished by both shape and color.

Meng: The takeaway for me is that the next focus needs to be on making sure this system can handle the computational load of running it efficiently on live video feeds without needing massive compute resources.

Lalam: I'm confident that this method provides a way for systems to automatically isolate specific networks like the RTP based on their visual signature, which is pretty concrete.

Tom: So, we've gone from just having an idea to having a system that can actually perform this clustering with solid numbers and clear guidance on how to tune those fusion parameters.

Jane: This paper gives us a solid blueprint for anyone trying to group similar objects without needing massive amounts of labeled data.

Lu: It’s a good step in building more flexible systems that can handle the ambiguity inherent in visual data very well.

Meng: We need to keep pushing on the scalability part because if it doesn't scale well, it won't be useful when we deploy this technology widely.

Lalam: The paper "Unsupervised Clustering of Public Transport Vehicles Using Deep Morphological Features and Color Histograms" gives us a clear way to automate the grouping of public transport vehicles based on their visual signature.

Conclusion: Tom: So we've been looking at "Unsupervised Clustering of Public Transport Vehicles Using Deep Morphological Features and Color Histograms." Basically, they figured out how to automatically group buses into types like different routes just by looking at their pictures, using a mix of deep learning features and color histograms.

Lu: This paper is important because it shows how deep learning features can be combined with traditional computer vision tasks in this specific way.

Meng: Next week, we're looking at something totally different, so keep your eyes peeled for that one.

Juan Irving Vasquez

Centro de Innovacion y Desarrollo Tecnolígico en CómputoŁ Instituto Politecnico Nacional

cs.CV

Submitted: 2026-10-15

Updated: 2026-10-15

Comments: Conference paper: under review

Code: https://github.com/ultralytics/ultralytics

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 72/100

The gist: The gist: This study presents an unsupervised clustering methodology for public transport vehicles that combines morphological features extracted with a frozen DINOv2 Vision Transformer and

Key concepts

Morphological Features (DINOv2): These are structural descriptions of vehicles extracted using a pre-trained Vision Transformer model called DINOv2. The model acts as a fixed feature extractor, capturing the vehicle's shape and structure without needing retraining. This provides a robust representation of the bus or vehicle's physical form.
HSV Color Histograms: These features quantify the color of vehicles by creating histograms across 32 bins for Hue and Saturation in the HSV color space. Using HSV separates chromatic information from brightness, and focusing on a central crop ensures these histograms primarily describe the vehicle's body color.
Weighted Early Fusion ($\alpha$): This technique combines the morphological features and color features into a single vector using a weighted average controlled by a parameter $\alpha$. $\alpha$ dictates how much influence the structural data versus the color data has on the final combined feature. Adjusting this weight allows researchers to find the optimal balance between shape and color.
UMAP Projection: Uniform Manifold Approximation and Projection (UMAP) is a dimensionality reduction technique used to map high-dimensional fused feature vectors into a lower-dimensional space. This step simplifies the data structure, making it easier for the subsequent K-Means clustering algorithm to effectively group similar vehicles based on their combined features.

Terminology

Clustering — 4× in this paper · explained in 2 episodes
K-means Clustering — 1× in this paper · explained in 3 episodes
Transformer — 1× in this paper · explained in 2 episodes

Summary

The gist: This study presents an unsupervised clustering methodology for public transport vehicles that combines morphological features extracted with a frozen DINOv2 Vision Transformer and handcrafted HSV color histograms through weighted early fusion, followed by UMAP projection and K-Means clustering.

Problem Addressed

The research addresses the challenge of grouping public transport vehicles captured by surveillance cameras into semantically meaningful categories, such as transit routes and operators, without using labeled data The problem is challenging for two complementary reasons. First, buses operating on different routes often share the same body structure and differ mainly in their paint schemes. Second, vehicles of different types, such as full-size buses, microbuses (small buses), and vans, may share similar colors while differing substantially in shape. A useful representation must therefore capture both morphology and color.

Methodology

The proposed pipeline involves a multistage approach where features are extracted first and clustered afterwards

Feature Extraction

Two complementary types of features are extracted from each image

** Morphological Features: These are extracted using DINOv2 [7], a family of Vision Transformer (ViT) [11] models trained through self-supervised learning. The model is used exclusively as a frozen feature extractor, meaning no fine-tuning or retraining was performed The color features are computed by calculating one-dimensional histograms with 32 bins for the Hue and Saturation channels in the HSV color space. The HSV space was preferred because it separates chromatic information from intensity A central crop covering 50% of the image area is applied before computing the histograms so that the descriptor predominantly reflects the vehicle body. The resulting color feature vector is formed by concatenating these histograms. The L2-normalized versions of both descriptors are then used for fusion This results in the fused vector vi. The fusion weight α controls the relative contribution of the morphological features and 1 − α that of the color features. This decomposition shows that α directly regulates the trade-off between structural and chromatic similarity. 4. Clustering and Evaluation The fused vectors are first projected onto a lower-dimensional manifold using Uniform Manifold Approximation and Projection (UMAP) [14] K-Means clustering [15] is then applied to these reduced representations This procedure is repeated independently for each fusion weight α in the set of values such as α ∈ 0.35, 0.50, and 0.65. The evaluation involves calculating accuracy, precision, recall, and F1-score by treating the cluster with the largest number of RTP buses as the predicted class The best configuration with α = 0.65 achieved the highest accuracy (0.9521), precision (0.9507), and F1-score (0.7516). The color information increased precision by 18 percentage points and the F1-score by 7.6 points compared to the morphological baseline The results show that a moderate color weight increases the precision of the target cluster with respect to morphological features alone<ref:2610.

Improvements for AI systems

Focus on unsupervised category-level partitioning of public transport vehicles, which has not been addressed in existing work because it groups different routes and vehicle types based on morphology and color. This allows systems to isolate specific transit networks like the Red de Transporte de Pasajeros (RTP) without requiring manual route labeling.
Implement a hybrid feature descriptor combining frozen DINOv2 morphological features and HSV color histograms through weighted early fusion, as described by the paper's structure: a morphological descriptor given by the class token of a frozen DINOv2 model, and a color descriptor given by hue and saturation histograms computed in the HSV color space.
Utilize UMAP projection to map the fused feature vectors before K-Means clustering, as this step is crucial for handling high-dimensional data: the fused vectors were first projected onto a lower-dimensional manifold using Uniform Manifold Approximation and Projection (UMAP).
Optimize the fusion weight alpha based on evaluation criteria to maximize cluster purity; for instance, setting alpha = 0.65 achieved the best overall performance, with the highest accuracy (0.9521), precision (0.9507), and F1-score (0.7516).
Develop a system capable of assessing the trade-off between structural and chromatic information by observing that a moderate color weight improves the purity of the target cluster, whereas over-prioritizing color degrades the structural understanding.
Enhance robustness against feature collapse by avoiding excessive reliance on one modality; specifically, ensure that color should complement, not replace, the morphological representation to prevent scenarios where over-prioritizing color degrades the structural understanding of the vehicles and leads either to over-segmentation or to the collapse of the clustering space.