VehicleNet-RFDETR-n is a multi-class vehicle detection model designed for fine-grained vehicle type recognition in real-world traffic scenes. It is fine-tuned on the UVH-26-MV Dataset, curated and released by the Indian Institute of Science (IISc), Bangalore, which captures the highly complex, dense, and heterogeneous nature of Indian road traffic.
The model recognizes 14 vehicle categories, including hatchbacks, sedans, SUVs, MUVs, two-wheelers, three-wheelers, buses, trucks, and a range of commercial vehicle types. This nano variant is optimized for low-latency inference, balancing speed and accuracy for deployment on resource-constrained hardware.
The model is fine-tuned on the RFDETRNano architecture (arXiv: 2511.09554) by Roboflow, using rfdetr version 1.6.1.
Model Specifications
Parameter
Value
Base Architecture
RFDETRNano
Number of Classes
14
Total Layers
-
Parameters
30.5 M
GFLOPs
-
Input Resolution
384 × 384
Training Epochs
8
Batch Size
4
Gradient Accumulation Steps
2
Effective Batch Size
16 (batch × grad_accum × GPUs)
Training Hardware
Dual NVIDIA Tesla T4 GPUs
Framework
Roboflow (PyTorch)
Pretrained Weights
RFDETRNano (Roboflow)
Performance Metrics
Metric
Value
mAP@50
0.66771
mAP@50:95
0.53883
mAP@75
0.59782
Precision
0.66409
Recall
0.63997
Training Curves
Training_Curves
Intended Use
VehicleNet-RFDETR-n8 is suitable for the following applications:
Traffic Surveillance & Analytics — Automated vehicle classification in urban and highway environments.
Edge Device Deployment — Optimized for low-latency inference on constrained hardware.
Academic Research & Benchmarking — Evaluation of fine-grained vehicle detection in heterogeneous traffic conditions, particularly on Indian road datasets.
Out-of-Scope Use
Deployment in safety-critical systems without independent validation.
Surveillance applications that violate individual privacy rights or applicable regulations.
Any use case inconsistent with the Apache License 2.0 terms.
Citation
If you use this model or the UVH-26-MV dataset in your research, please cite the respective dataset and model sources appropriately.
License
This model is released under the Apache License 2.0. You are free to use, modify, and distribute this model subject to the terms of the license. See the LICENSE file for full details.
