CSP

A pytorch implementation of CSP, Wei Liu, etc. High-level Semantic Feature Detection: A New Perspective for Pedestrian Detection, CVPR 2019.

The authors' Keras implementation liuwei16/CSP.

Updates

• 20 March 2020, Load keras weights into pytorch model, and got the same model outputs as resnet50 nn_p3p4p5 in liuwei16/CSP/resnet50.py
• 23 March 2020, finish test_caltech, got 4.06 MR under Reasonable setting using authors trained weights.
• 24 March 2020, finish test_city, got 11.02 MR under Reasonable setting using authors trained weights.
• 27 March 2020, finish test_one_pic_face, load wider face trained model and test face detection on one pic.

Environments

• ubuntu16.04
• cuda10.0
• python3.6
• pytorch1.4.0

Test One Picture Pedestrian

Download trained weights from Wei Liu's baidu netdisk BaiduYun (Code: jcgd).

For this test, we are using ResNet-50 initialized from CityPersons: Height+Offset prediciton: model_CSP/caltech/fromcity/net_e82_l0.00850005054218.hdf5

We'll use this keras hdf5 weights, and load it in pytorch, and run inference.

git clone https://github.com/wang-xinyu/csp.pytorch
cd csp.pytorch
// put net_e82_l0.00850005054218.hdf5 here
python test_one_pic_pedestrian.py

The out.jpg as follows will be generated.

Eval Caltech

We still use net_e82_l0.00850005054218.hdf5 above.

Put caltech test images(4024 images) in data/caltech/images/. The dir tree is as below.

data
    caltech
        test.txt
        images
            set06_V000_I00029.jpg
            set06_V000_I00059.jpg
            ...
            set10_V011_I01709.jpg

Get AS.mat from liuwei16/CSP, put into eval_caltech/

Then run test_caltech.py to generate results, and use octave to run eval and get MR.

python test_caltech.py
cd eval_caltech/
git clone https://github.com/pdollar/toolbox.git
sudo apt install octave
octave-cli
   -> dbEval

A result file eval_caltech/ResultsEval/eval-newReasonable.txt will generated.

Eval CityPersons

For this test, we are using Height+Offset prediction: model_CSP/cityperson/withoffset/net_e121_l0.hdf5, download it from the baidu link above.

Put citypersons images in data/citypersons/. The dir tree is as below.

data
    citypersons
        leftImg8bit
            train
                aachen
                bochum
                ...
                zurich
            val
                frankfurt
                lindau
                munster

Then run test_city.py to generate results, and run eval and get MR.

python test_city.py
cd eval_city/
python dt_txt2json.py
cd eval_script/
python eval_demo.py

Test One Picture Face

Download trained weights from Wei Liu's baidu netdisk BaiduYun (Code: jcgd).

For this test, we are using ResNet-50 Height+Width+Offset prediction: model_CSP/widerface/net_e382_l0.hdf5 We'll use this keras hdf5 weights, and load it in pytorch, and run inference.

cd csp.pytorch
// put net_e382_l0.hdf5 here
wget https://github.com/TencentYoutuResearch/FaceDetection-DSFD/raw/master/data/worlds-largest-selfie.jpg
python test_one_pic_face.py

The out.jpg as follows will be generated.