Overview
MobileNetV2-SSD, originally described in MobileNetV2: Inverted Residuals and Linear Bottlenecks published on January 13, 2018, and SSD: Single Shot MultiBox Detector published on December 8, 2015, pairs a lightweight MobileNetV2 as its feature extractor with an SSD head for predicting bounding boxes and classes, designed to run real-time object detection on mobile and embedded devices. The model takes a single RGB image of 300x300 pixels as input and outputs bounding box coordinates and class scores. This repository contains a pre-compiled model optimized for NXP Ara240 Discrete NPU, making it suitable for edge computing scenarios requiring efficient inference with minimal latency.
Key Features
- The SSD head predicts bounding boxes and class probabilities for multiple objects simultaneously in one forward pass, enabling real-time inference.
- Additional convolutional feature layers appended to the base network produce a fixed set of detection predictions at different scales using a set of convolutional filters.
- The lightweight architecture of the MobileNetV2-SSD reduces computational complexity while maintaining accuracy for object detection tasks.
- Depthwise separable convolutions significantly reduce parameter count and the number of operations, making the model well suited for embedded systems.
Model Description
This is a quantized and compiled version of MobileNetV2-SSD optimized for Ara240 DNPU.
- Base Model: ssd_mobilenet_v2_coco
- Original Model Authors: TensorFlow Team
- Original License: Apache-2.0
- Modified by: NXP
Modifications
This model is a derivative work with the following changes from the original:
- Quantization: INT8 calibrated using COCO val2017
- Compilation: Compiled for Ara240 DNPU
- Format: Converted to DVM format for NPU deployment via the Ara SDK
Original model available at: MobileNetV2-SSD. For more information on quantization, compilation and evaluation procedures for this model, refer to its dedicated Ara-Model-Zoo page: mobilenetv2-ssd. There you will find the steps and recipes used to quantize, compile and evaluate this model for Ara240 DNPU deployment.
Model Information
| Information | Value |
|---|---|
| Input shape | RGB image [1, 3, 300, 300], int8 |
| Input example | (Image source: Wikimedia Commons, Public Domain) |
| Output shape | [1, 1917, 4] - Bounding box coordinates [1, 1917, 91] - Class confidence scores |
| Output example | ![]() |
| Number of parameters | 16.9 M |
| File size | 19 MB |
Supported Hardware and Software Configurations
| Model binary | Ara SDK | Precision | Link to model binary |
|---|---|---|---|
| mobilenetv2-ssd.dvm | r3.0 | INT8 | Download |
Limitations and Biases
This model inherits all limitations from the original MobileNetV2-SSD model. Additional limitations:
- Hardware-specific: Only runs on hardware with an Ara240 DNPU; not compatible with standard CPU/GPU inference frameworks.
- Quantization effects: INT8 quantization may introduce accuracy differences relative to the FP32 baseline.
Getting the Model
This repository uses git tags to version the pre-compiled DVM file by Ara SDK release. Use the Hugging Face CLI to download:
# Download the model for a specific Ara SDK release (e.g., r3.0)
hf download nxp/MobileNetV2-SSD-Ara240 mobilenetv2-ssd.dvm --revision r3.0 --local-dir .
The --local-dir . flag saves the files directly to the current directory instead of
the default Hugging Face cache (~/.cache/huggingface/hub).
License
This model is released under the Apache-2.0 License, the same license as the original MobileNetV2-SSD model.
Citation
IEEE Format:
M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, "MobileNetV2: Inverted Residuals and Linear Bottlenecks," arXiv preprint arXiv:1801.04381, 2019. [Online]. Available: https://arxiv.org/abs/1801.04381.
W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and A. C. Berg, "SSD: Single Shot MultiBox Detector," in Proc. European Conf. Comput. Vis., 2016, pp. 21β37. [Online]. Available: https://arxiv.org/abs/1512.02325.
H. Yu, C. Chen, X. Du, Y. Li, A. Rashwan, L. Hou, P. Jin, F. Yang, F. Liu, J. Kim, and J. Li, "TensorFlow Model Garden," 2020. [Online]. Available: https://github.com/tensorflow/models.
BibTeX Format:
If you use this model, please cite both this work and the original model:
@misc{sandler2019mobilenetv2invertedresidualslinear,
title={MobileNetV2: Inverted Residuals and Linear Bottlenecks},
author={Mark Sandler and Andrew Howard and Menglong Zhu and Andrey Zhmoginov and Liang-Chieh Chen},
year={2019},
eprint={1801.04381},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/1801.04381},
}
@inbook{Liu_2016,
title={SSD: Single Shot MultiBox Detector},
ISBN={9783319464480},
ISSN={1611-3349},
url={http://dx.doi.org/10.1007/978-3-319-46448-0_2},
DOI={10.1007/978-3-319-46448-0_2},
booktitle={Computer Vision β ECCV 2016},
publisher={Springer International Publishing},
author={Liu, Wei and Anguelov, Dragomir and Erhan, Dumitru and Szegedy, Christian and Reed, Scott and Fu, Cheng-Yang and Berg, Alexander C.},
year={2016},
pages={21β37}
}
@misc{tensorflowmodelgarden2020,
author = {Yu, Hongkun and Chen, Chen and Du, Xianzhi and Li, Yeqing and Rashwan, Abdullah and Hou, Le and Jin, Pengchong and Yang, Fan and Liu, Frederick and Kim, Jaeyoun and Li, Jing},
title = {{TensorFlow Model Garden}},
howpublished = {\url{https://github.com/tensorflow/models}},
year = {2020}
}
