---
license: agpl-3.0
extra_gated_prompt: >-
### UrbanFlow Intelligence Engine | Model Access & Usage Agreement
Access to the UrbanFlow perception architectures is governed by this professional agreement.
By requesting access, downloading, or utilizing these models, you confirm your commitment to the
following terms and our open-source licensing structure:
**Licensing & Attribution**: UrbanFlow is built upon the foundational advancements of the
Ultralytics YOLO architectures. In alignment with the **AGPL-3.0 license**, we release this
fine-tuned engine to the community while strictly respecting the open-source contributions of
**Ultralytics** and their respective authors. We acknowledge their pivotal role in advancing
the boundaries of real-time object detection.
1. **Legal Capacity**: You certify that you meet the legal age of majority in your
jurisdiction and possess the authority to accept and comply with these terms.
2. **Intent of Usage**: UrbanFlow is provided for technical evaluation, academic mobility
research, and urban planning analysis. In accordance with the AGPL-3.0 "copyleft"
provisions, any derivative works or hosted services utilizing this model must also
be made available under the same license.
3. **No Warranty for Critical Infrastructure**: This model is provided "as-is" for research
and evaluation purposes. Perception365 makes no guarantees regarding absolute accuracy in
safety-critical autonomous navigation or high-stakes regulatory environments. Independent
validation is mandatory for any production-grade deployment.
4. **Operational Accountability**: You assume sole responsibility for the deployment and
outputs of the model. Usage for unlawful surveillance or any application violating
individual privacy standards is strictly prohibited.
If you do not agree to these professional standards or the AGPL-3.0 licensing terms,
do not proceed with this access request.
extra_gated_fields:
Full Name: text
Organization or Institution: text
Work or Student Email: text
Country: country
Professional or Academic Role:
type: select
options:
- Undergraduate / Graduate Student
- Academic Researcher / Professor
- Computer Vision / AI Engineer
- Traffic / Urban Planning Engineer
- Other Professional / Consultant
Primary Use Case:
type: select
options:
- Academic Research & Publication
- Model Benchmarking & Evaluation
- Smart City Pilot / Deployment
- Heterogeneous Traffic Analysis
- Industrial Monitoring & Testing
I certify that I have read and agree to the UrbanFlow Usage Agreement and AGPL License: checkbox
extra_gated_button_content: Request Access
datasets:
- iisc-aim/UVH-26
language:
- en
metrics:
- confusion_matrix
library_name: ultralytics
base_model:
- Ultralytics/YOLO26
pipeline_tag: object-detection
tags:
- indian-traffic
- inference-efficiency
- multi-vehicle-detection
- ultralytics
- edge-computing
---
# VehicleNet-Y26s
**VehicleNet-Y26s** is another multi-class vehicle detection model designed for fine-grained vehicle type recognition in real-world traffic scenes. The model is trained on **`UVH-26-MV Dataset`** released by IISc Banaglore. The dataset is based on Indian traffic which is highly challenging, dense and heterogeneous. It contains 14 vehicle categories such as hatchback, sedan, SUV, MUV, two-wheelers, three-wheelers, buses, trucks, and commercial vehicles.
This `s` variant is designed for speed and inferences on low-latency devices, offering significant speed and accuracy.
This model is finetuned on `YOLO26s`:[arXiv](https://arxiv.org/html/2509.25164v3) model by **`Ultralytics`** using **`UVH-26-MV Dataset`**.
## Model Overview and Parameters
- Pretrained_weights: YOLO26s
- Number of Classes: 14
- Layers: 122 layers
- Parameters(M): 9,470,598 parameters, 0 gradients
- GFLOPs: 20.6
- Input Resolution: 640 × 640
- Training Epochs: Up to 60 (early stopping applied, patience=5), best model at: 40/60
- Batch Size: 80
- Hardware: Dual NVIDIA Tesla T4 GPUs
- Framework: Ultralytics YOLO (PyTorch)
## Performance Summary
- `mAP@50`: 0.7274
- `mAP@50:95`: 0.64294
- `Precision`: 0.68078
- `Recall`: 0.69005
## Per-class mAP@50:95
The model showed strong detection performance for structurally distinct vehicle categories such as two-wheelers, three-wheelers, buses, and trucks. Fine-grained car subclasses (hatchback, sedan, SUV, MUV) exhibit expected inter-class confusion/challenge due to visual similarity and viewpoint overlap, as reflected in the confusion matrix.
## INT8 OpenVINO Weights
This repository now includes **INT8-quantized OpenVINO weights** for efficient CPU inference.
The model was exported using the following configuration:
```python
model.export(
format="openvino",
imgsz=736,
dynamic=False,
int8=True,
device='cpu',
batch=2,
data="uvh-26-mv-y26/UVH_Y26/data.yaml",
fraction=0.25
)
```
Checkout at:[best_int8_openvino_model](https://huggingface.co/Perception365/VehicleNet-Y26s/tree/main/weights/best_int8_openvino_model)
## Intended Use
The model is suitable for:
- Edge device computation
- Traffic surveillance and analytics
- Academic research and benchmarking
## License
This model is released under the `Apache License 2.0`.