TRAINING AN EMBEDDED OBJECT DETECTOR FOR INDUSTRIAL SETTINGS WITHOUT REAL IMAGES

Table of content

• Overview
• Installation
• Dataset
• Evaluation
• Training
• Citation
• Acknowledgements

Overview

This repository contains the code for MobileNet-V3 SSD, MobileNet-V3 Small SSD, and MobileNet-V2 SSD as described in the paper. The weights of the models trained on the T-LESS dataset are available for download, as well as the compared Mask R-CNN model.

Installation

Project

Clone the current repository:

git clone --recurse-submodules https://gitlab.liris.cnrs.fr/jcohen/synthetic-ssd.git

The "Object Detection Metrics" project should be downloaded in the deps folder. After download, move the lib subfolder and _init_paths.py file into synthetic_ssd\metrics.

Environment

Create and activate a conda environment.

conda create --file requirements.yml

This project has been tested on Windows with Python 3.8 and PyTorch 1.9, but previous versions of Python 3 and PyTorch should also work.

Dataset

The T-LESS dataset is available on the BOP website. We use the "PBR-BlenderProc4BOP training images" and "All test images" subsets, saved in the data folder. The expected folder structure is:

- data
| - tless
  | - test_primesense
  | - train_pbr

Otherwise, you can change the TLESS_BASE_PATH variable in synthetic_ssd\config.py

Evaluation

For the Mask R-CNN model, download the trained weights file from the CosyPose project into the data/weights folder, or elsewhere and change the MASK_RCNN_PATH variable in synthetic_ssd\config.py. The model evaluated in the paper corresponds to the one with model_id detector-bop-tless-pbr--873074.

To reproduce the performance reported in the paper, use the evaluation script:

python -m synthetic_ssd.scripts.run_eval \
--model [mobilenet_v2_ssd, mobilenet_v3_ssd, mobilenet_v3_small_ssd, mask_rcnn] \
--tf [aug1, aug2]

Default values are set to mobilenet_v2_ssd and aug2.

Training

Coming soon...

Citation

If you use this code in your research, please cite the paper:

@inproceedings{cohen2021training,
  title={Training An Embedded Object Detector For Industrial Settings Without Real Images},
  author={Cohen, Julia and Crispim-Junior, Carlos and Chiappa, Jean-Marc and Tougne, Laure},
  booktitle={2021 IEEE International Conference on Image Processing (ICIP)},
  pages={714--718},
  year={2021},
  organization={IEEE}
}

Acknowledgements

This repository is based on the following works:

• MobileNet-V3 with ImageNet pretraining: https://github.com/d-li14/mobilenetv3.pytorch
• SSD: https://github.com/sgrvinod/a-PyTorch-Tutorial-to-Object-Detection
• MobileNet-V2 - SSD: https://github.com/qfgaohao/pytorch-ssd
• MobileNet-V3 - SSD: https://github.com/tongyuhome/MobileNetV3-SSD

This work is supported by grant CIFRE n.2018/0872 from ANRT.