brunosauvalle/bb-sgd

Original implementation of the background reconstruction model described in the paper "Fast and accurate background reconstruction using background bootstrapping"

BB-SGD : background boostrapping using stochastic gradient descent

This repository is the official implementation of the model BB-SGD described in the paper "BB-SGD : fast and accurate background reconstruction using background bootstrapping"

Requirements

The model requires Pytorch (>= 1.7.1) and Torchvision with cuda capability

The model also requires OpenCV (>=4.1)

To install other requirements:

pip install -r requirements.txt

The model has been tested with Nvidia RTX 2080 TI and Nvidia RTX 3090 GPU.

How to use the model

the command to generate the background from a sequence of frames is

python main.py --input_path your_input_path 

where your_input_path is the path to the folder where the frame sequence is saved. Example : python main.py --input_path /workspace/Datasets/SBMnet_dataset/basic/511/input

the result background image will be stored in the current working directory. To view options, type python main.py -h

Evaluation

To evaluate the BB-SGD model on the SBMnet 2016 dataset:

• download the SBMnet 2016 dataset from the following link :

 http://pione.dinf.usherbrooke.ca/static/dataset/SBMnet_dataset.zip

and save it to some folder

• generate the backgrounds using the command

python SBM_processing_pipeline.py datasetPath resultPath

where datasetPath is the path to the saved SBMnet dataset and resultPath is the path to the result folder where you want to store the results.

Example : python SBM_processing_pipeline.py /workspace/Datasets/SBMnet_dataset /workspace/Datasets/SBMnet_results

This pipeline will compute and save the 79 backgrounds associated to the 79 sequences of the dataset

In order to compute statistics for the sequences where a ground truth is provided with the SBMnet dataset, use the command

python SBM_UTILITY.py groundtruthtPath resultPath

example : python SBM_UTILITY.py /workspace/Datasets/SBMnet_dataset /workspace/Datasets/SBMnet_results the evaluation statistics will be stored as cm.txt file in the resultPath folder

If you want also to complete this evaluation with other sequences where a ground truth background is also publicly available but not provided with the SBMnet dataset, go the website https://sbmi2015.na.icar.cnr.it/SBIdataset.html, download the groundtruth images associated to the sequences Toscana, Candela_m1.10,CaVignal,Foliage, People&Foliage, and add them to the appropriate folder of the SBM dataset before starting the evaluation utility.

Warning : Different runs of the model with the same inputs may lead to small differences in evaluation results ompared to the results due to the random initialization of the image.

Acknowledgment

The processing and evaluation codes for the SBMnet dataset ( SBM_Evaluate.py, SBM_gauss.py, SBM_processing_pipeline.py and SBM_UTILITY.py) are adapted from the evaluation codes available on the SBMnet webiste (links http://pione.dinf.usherbrooke.ca/code)