Pytorch implementation of Bi-box Regression as described in Bi-box Regression for Pedestrian Detection and Occlusion Estimation (ECCV2018).

The implementation is not just the same as the paper. We use one-stage detector RetinaNet followed by this repo. And some details may be different.

• Python3
• Pytorch0.4
• torchvision
• tensorboardX

Install packages.

sudo apt-get install tk-dev python-tk
pip install cffi
pip install cython
pip install pandas
pip install tensorboardX

Build NMS.

cd Bi-box_Regression/lib
sh build.sh

Create folders.

cd Bi-box_Regression
mkdir ckpt mAP_txt summary weight

This repo is built for human detection. The popular annotation format for human detection（or pedestrian detection) contains three parts: full bbox, visible bbox and ignore regions such as Citypersons and Crowdhuman. You should write them in CSV or TXT files.

Three examples are as follows:

$image_path/img_1.jpg x1 y1 x2 y2 x1v y1v x2v y2v person
$image_path/img_1.jpg x1 y1 x2 y2 ignore
$image_path/img_2.jpg . . . . .

x1 y1 x2 y2 mean full bbox and x1v y1v x2v y2v mean visible bbox. Images with more than one bounding box should use one row per box. Labels that we often use are 'person' or 'ignore'. When an image does not contain any bounding box, set them '.'.

A TXT file (classes.txt) is needed to map label to ID. Each line means one label name and its ID. One example is as follows:

person 0

We use resnet18, 34, 50, 101, 152 as the backbone. You should download them and put them to /weight.

• resnet18: https://download.pytorch.org/models/resnet18-5c106cde.pth
• resnet34: https://download.pytorch.org/models/resnet34-333f7ec4.pth
• resnet50: https://download.pytorch.org/models/resnet50-19c8e357.pth
• resnet101: https://download.pytorch.org/models/resnet101-5d3b4d8f.pth
• resnet152: https://download.pytorch.org/models/resnet152-b121ed2d.pth

python train.py --csv_train <$path/train.txt> --csv_val <$path/val.txt> --csv_classes <$path/classes.txt> --depth <50> --pretrained resnet50-19c8e357.pth --model_name <model name to save>

• Focal Loss for Dense Object Detection
• Bi-box Regression for Pedestrian Detection and Occlusion Estimation