Quantization-Aware Training for Efficient On-Board Object Detection on FPGAs: Case Studies

This paper proposes a Quantization-Aware Training (QAT) framework enhanced with object-scale-aware regularization to mitigate accuracy degradation in small-object detection caused by quantization noise, specifically targeting resource-constrained FPGA deployments for geoscience and remote sensing applications.

Case studies

Case Study 1: Bird Detection at Airports

A 3D-printed case is designed to accommodate the Xilinx Kria KV260 FPGA with waterproof sealing, cable management, and air cooling. The board connects to an IP camera for video streaming through the RTSP protocol.

Some birds have been detected at Airport Oberpfaffenhofen, Weßling, Germany

Case Study 2: Aerial-View Building Detection

Visualization of the building detection task with true labels (above) and detected building footprints (below). Several small cabins in white rectangles are newly detected and absent from the dataset (BBD).

Usage

Environment

1. Docker env

   docker run -e UID=$(id -u) -e GID=$(id -g) --name qatdet --gpus device=0 -d -it --shm-size 32G --mount source=$(pwd),target=/workspace,type=bind tumbgd/vai-pt-cuda

   docker exec -it qatdet bash

2. Inside docker container qatdet

   python -m pip install --user -r requirements.txt
   cd code
   python -m pip install --user -v -e .
   cd ..

   In this way, the yolox library is installed. The installation is successful if you have the following output:

   Installed /workspace/code
   Successfully installed yolox
   

   All following steps shall be executed in this docker environment until we obtain a compiled .xmodel.

Dataset preparation

Download the datasets and transform them into COCO format.

Bavaria Building Dataset (BBD)

Download the dataset from here. We use bbd2k5-images-image.tar.bz2 and bbd2k5-images-umring.tar.bz2 in this project. Unzip them and put them into ./bbd/data. The resulting directory tree should be like

.
|-- LICENSE
|-- README.md
`-- bbd
    `-- data
        |-- bbd2k5-images-image
        `-- bbd2k5-images-umring

Then we can generate COCO-format json files by

python ./bbd/Mask2COCO.py

For detailed dataset description, please check here.

dataset citation

@inproceedings{10.1145/3589132.3625658,
    author = {Werner, Martin and Li, Hao and Zollner, Johann Maximilian and Teuscher, Balthasar and Deuser, Fabian},
    title = {Bavaria Buildings - A Novel Dataset for Building Footprint Extraction, Instance Segmentation, and Data Quality Estimation},
    year = {2023},
    isbn = {9798400701689},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3589132.3625658},
    doi = {10.1145/3589132.3625658},
    booktitle = {Proceedings of the 31st ACM International Conference on Advances in Geographic Information Systems},
    articleno = {108},
    numpages = {4},
    location = {Hamburg, Germany},
    series = {SIGSPATIAL '23}
}

MVA2023: Small Object Detection Challenge for Spotting Birds

You can find the download link in the dataset repository here. We only use the drone2021 (~63 GB) part of this dataset. Put unzipped files into bird/data. The resulting directory tree should be like

|-- LICENSE
|-- README.md
`-- bird
    |-- data
    |   |-- annotations
    |   |   |
    |   |   `--.json
    |   `-- images
    |       |--1
    |       |   `--.jpg
    |       |--2
    |       `--...
    `-- usable_images_updater.py

The annotations given in the original dataset is already COCO-format. We use the following script to remove images containing no birds or birds that are too small (< 40 pxl width for 4K images).

python bird/Filter.py

dataset citation

@inproceedings{mva2023_sod_challenge,
  title={{MVA2023 Small Object Detection Challenge for Spotting Birds: Dataset, Methods, and Results}},
  author={Yuki Kondo and Norimichi Ukita and Takayuki Yamaguchi and Hao-Yu Hou and Mu-Yi Shen and Chia-Chi Hsu and En-Ming Huang and Yu-Chen Huang and Yu-Cheng Xia and Chien-Yao Wang and Chun-Yi Lee and Da Huo and Marc A. Kastner and Tingwei Liu and Yasutomo Kawanishi and Takatsugu Hirayama and Takahiro Komamizu and Ichiro Ide and Yosuke Shinya and Xinyao Liu and Guang Liang and Syusuke Yasui},
  booktitle={2023 18th International Conference on Machine Vision and Applications (MVA)},
  note={\url{https://www.mva-org.jp/mva2023/challenge}},
  year={2023}
}

Training

Perform QAT and get .xmodel for AMD Kria KV260 FPGAs.

• For BBD:

  bash code/bbd.sh -n --osa  # bbd YOLOX-nano w/  osa
  bash code/bbd.sh -n        # bbd YOLOX-nano w/o osa
  bash code/bbd.sh -t --osa  # bbd YOLOX-tiny w/  osa
  bash code/bbd.sh -t        # bbd YOLOX-tiny w/o osa

• For MVA2023:

  bash code/bird.sh -n --osa  # bird YOLOX-nano w/  osa
  bash code/bird.sh -n        # bird YOLOX-nano w/o osa
  bash code/bird.sh -t --osa  # bird YOLOX-tiny w/  osa
  bash code/bird.sh -t        # bird YOLOX-tiny w/o osa

  After training, you should find YOLOX_0_int.xmodel at ./YOLOX_outputs/{bird | bbd}/convert_qat_results. You can also change the output directory by modifying the script specified by CFG.

Compilation for FPGA

The target deployment platform in this project is AMD (Xilinx) Kria KV260 FPGAs. After QAT, we compile the model obtained in the previous step for KV260.

vai_c_xir -x <PATH_TO/YOUR.xmodel> -a /opt/vitis_ai/compiler/arch/DPUCZDX8G/KV260/arch.json -o <EXPORT_PATH> -n <NEWNAME>

If you are playing with another kind of board, you need to change the value for the -a option. Moreover, you can export the computation graph to default.svg of xmodel by

xdputil xmodel <PATH_TO_COMPILED.xmodel> -s

You can find more information in the official document here.

On-board Deployment

Switch to the README in onboard, which also contains trained onboard model for reproduction.

Citation

@Article{isprs-annals-XI-2-2026-697-2026,
  AUTHOR={Luo, X. and Fogarasi, G. and Syrgak, A. and Walther, P. and Werner, M.},
  TITLE={{Quantization-Aware Training for Efficient Object Detection on FPGAs: Case Studies}},
  JOURNAL={ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences},
  VOLUME={XI-2-2026},
  YEAR={2026},
  PAGES={697--704},
  URL={https://isprs-annals.copernicus.org/articles/XI-2-2026/697/2026/},
  DOI={10.5194/isprs-annals-XI-2-2026-697-2026}
}

Acknowledgement

This repository is built based on pt_yolox-nano_3.5 in the Model Zoo of Vitis AI. Keep an eye on COCO path specifications in ./code/yolox/exp/yolox_base.py. YOLOX-tiny variants are also built upon this.

Copyright 2022-2023 Advanced Micro Devices Inc.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

See also YOLOX.

License

Licenced under MPL-2.0 license (LICENSE or https://opensource.org/license/mpl-2-0).