This repository contains a software framework designed for experimentation with different classical and state-of-the-art pedestrian detection algorithms and benchamrk datasets.
The code is written in Matlab and contains modified scripts based on the original code published by the authors of the algorithms. There are utility functions based on the Piotr's toolbox allowing one to extract and manipulate data and evaluate the performance of the algorithms.
- Piotr's toolbox
- Caltech evaluation/labeling
Download and copy the above files to utilities folder.
PBF works with 8 pedestrian detection datasets including JAAD, CityPersons, Caltech, KITTI, Inria, ETH, TUD-Brussels, and Daimler.
The API uses the standardized versions of ETH, Inria, TUD-Brussels and Daimler in .seq and .vbb format which can be downloaded from here. For CityPersons, PBF uses the annotations in .mat format from here. For other datasets, download the data from the correspoding links.
The datasets should be placed in a common folder, as follows:
<main_folder>
/<dataset>
/<annotations_folder>
/<videos_or_images_folder>
For datasets Caltech, ETH, Inria, TUD-Brussels and Daimler copy the content of annotations.zip to annotations folder and copy set##.tar to videos/set## folders. In the case of JAAD dataset, copy vbb_full and vbb_part folders to annotations folder and all .seq files to video/set00 folder.
Copy the annotations of CityPersons to annotations folder and keep the images folder name as is. For KITTI dataset, just copy and extract the data and annotation folders as they are to the dataset folder.
PBF currently implements 10 classical and state-of-the-art pedestrian detection algorithms including ACF, CCF, Checkerboards, DeepPed, LDCF++, Faster-rcnn, MS-CNN, RPN+BF, SDS-RCNN, SPP.
For more detail regarding the selection of parameters for algorithms consult with the corresponding publications.
To switch between the models with dependency on Caffe, Matlab environment has to be restarted each time.
The algorithm is tested using cuda 8.
For compatibality with new changes in caffe do the following changes:
In cnnPyramid.m: Comment out line 13, %addpath(['path_to_caffe_codes' '/matlab/caffe']); and add parameter test to line 113 as in caffe('init', cnn.model_def, cnn.model_file, 'test');.
Comment out line 114 % caffe('set_phase_test');.
Download the pretrained model VGG_ILSVRC_16_layers.caffemodel as instructed by the authors and copy to data/CaffeNets.
The code is tested with Checkerboards_CVPR15_codebase version of the code.
Only the detection code is available for this algorithm.
The code is tested with cuda 7.
This algorithm depends on Piotr's toolbox. Make sure the toolbox in utilities folder is in Matlab's path.
The code is tested with cuda 7.
For the purpose of pedestrian detection replace the models in /models with /jaad/models_fcnn_ped.
The code is tested with cuda 8.
Download the pretrained VGG weights as intructed by the authors and place in a folder called. ms-cnn/pretrained
Downlowad the pretrained model as instructed by the authors and place in a folder called sds-rcnn/pretrained.
Only the detection code is available.
The code relies on optical flow, therefore the performance is optimal with sequential datasets such as JAAD and Caltech (with small skips).
To use the PBF whether for training and testing or data extraction, simply use experimentScript.m script.
PFB can be used to extract data for training/testing purposes with certain characteristics, e.g. scale of the images, type of annotations, or which portion of the data to use. For example the following example extracts partial occlusion annotations from mix subset of JAAD with skip of 10, 30, 30 for training, testinng and validation data in 0.5 scale.
dataInfo.dataSet = 'jaad';
dataInfo.anotationType = 'part';
dataInfo.jaadSubDatasetDir = 'data_indices';
dataInfo.jaadSubDataset = 'mix';
dataInfo.skipTrain = 10;
dataInfo.skipTest = 30;
dataInfo.skipVal = 30;
dataInfo.scale = 0.5;
dataInfo.justDataExtract = 1;
Note that setting parameter dataInfo.justDataExtract to 1 make the script to only extract the data.
The extracted data is saved under pedestrian_benchmark/data/<dataset_name>/<data_subset>. The training data folder is appended with the value of skip, e.g. in the above example the folder will be jaad_mix_scale_0.5/train10/. This allows the user to have multiple trining data with different sizes. For testing and validation data, the folder name appears only as test and val, so if new changes in testing or validation data is needed, the corresponding data folders should be deleted manually.
For training set up dataset's parameters as desired (see above) and set opts.trOrDet to either 0 (only training) or 1 (training followed by testing). If the dataset with desired parameters exists in /data folder, the algorithm starts training, and if not, the data first will be extracted followed by training.
Next, the algorithm parameters for training should be setup accordingly. The following example sets up ms-cnn algorithm to be trained using bounding boxes in the range of [50 inf], ignoring the bounding boxes with occlusion tag, and input training images of size 720x1280.
modelSetup = setAlgorithmParams('ms-cnn', 'modelName', 'ms-cnn-example', 'hRng', [50 inf], 'ignOcc', 1, ...
'scales', 720, 'max_size', 1280);
For more detail on how to set algorithms parameters check the comments in setAlgorithmParams.m.
The data set up for testing is similar to training. Simply specify what dataset and with which parameters are desired to run. To perform testing, set opts.trOrDet to either 1 or 2 (only testing). The following example runs testing for ms-cnn algorithm:
dataInfo.testImgsRange = []; % test on all testing data
opts.trOrDet = 2; % only testing
modelSetup = setAlgorithmParams('ms-cnn', 'modelName', 'ms-cnn-example',
'test_scales', 720, 'test_max_size', 1280);
Note that if the opts.trOrDet = 1 and model name specified under 'modelName' already exists, the training stage is skipped.
The detection results will be saved under detection_results/<algorithm_name>/<model_name>/<dataset_name>/<some_name>_dets.txt. Note that if the detection file already exists, i.e. the chosen model has been tested on a given dataset, testing stage will skip. If another round of testing is needed, set 'reapply' parameter in setAlgorithmParams to 1.
To perform evaluation after testing set opts.doEvaluation to true. Similar to algorithms, evaluation parameters can be set using setEvaluationParams function. The following example performs evalution by fixing the aspect ratio of ground truth boxes to 0.41, horizontal limit to [50 inf], visibility range to [0.65 1] with overlap threshold of 0.5 and saves the miss rate results as a text file.
evalSetup = setEvaluationParams('hRng',[50 inf], 'vRng', [.65 1], 'threshold', 0.5, 'saveMR', 1);
The result of evaluation is reported in terms of MR2, MR4 and mean IOU.
If as part of your research, you use any of the algorithms or datasets in PBF, cite the corresponding publications as instructed by the authors. If you use PBF framework for experimentation and benchmarking, please cite our paper:
@inproceedings{rasouli2018role,
title={It’s Not All About Size: On the Role of Data Properties in Pedestrian Detection},
author={Rasouli, Amir and Kotseruba, Iuliia and Tsotsos, John K},
booktitle={ECCVW},
year={2018}
}
and Piotr's toolbox for dependencies.
This is a work in progress with the aim of easing up future experimentation and benchmarking in the field of pedestrian detection. If you have a published work with a code written in Matlab, or a large scale dataset that want to be included in this framwork, please contact aras@eecs.yorku.ca.