Single Shot Multi-Box Detect Infer & Repeat implementation in PyTorch
Requirements:
- Install
poetry(https://python-poetry.org/docs/#installation) - Use
poetryto handle requirements- Execute
poetry add <package_name>to add new library - Execute
poetry installto create virtualenv and install packages
- Execute
To train the model use the train.py script. Activate the environment by running poetry shell and run python train.py --help to see all the available options. See supplementary material for details on hyperparameter settings in the research.
SSDIR re-uses trained SSD models and shares same datasets implementation. See the repository for more details.
Use make to run commands
make help- show helpmake test- run testsargs="--lf" make test- run pytest tests with different arguments
make shell- run poetry shell
In supplementary materials we shared more details about the model.
If you use SSDIR in your research, please consider citing:
@InProceedings{10.1007/978-3-031-08757-8_51,
author="Zieli{\'{n}}ski, Piotr
and Kajdanowicz, Tomasz",
editor="Groen, Derek
and de Mulatier, Cl{\'e}lia
and Paszynski, Maciej
and Krzhizhanovskaya, Valeria V.
and Dongarra, Jack J.
and Sloot, Peter M. A.",
title="Learning Scale-Invariant Object Representations with a Single-Shot Convolutional Generative Model",
booktitle="Computational Science -- ICCS 2022",
year="2022",
publisher="Springer International Publishing",
address="Cham",
pages="613--626",
abstract="Contemporary machine learning literature highlights learning object-centric image representations' benefits, i.e. interpretability, and the improved generalization performance. In the current work, we develop a neural network architecture that effectively addresses the task of multi-object representation learning in scenes containing multiple objects of varying types and sizes. In particular, we combine SPAIR and SPACE ideas, which do not scale well to such complex images, and blend them with recent developments in single-shot object detection. The method overcomes the limitations of fixed-scale glimpses' processing by learning representations using a feature pyramid-based approach, allowing more feasible parallelization than all other state-of-the-art methods. Moreover, the method can focus on learning representations of only a selected subset of types of objects coexisting in scenes. Through a series of experiments, we demonstrate the superior performance of our architecture over SPAIR and SPACE, especially in terms of latent representation and inferring on images with objects of varying sizes.",
isbn="978-3-031-08757-8"
}