rainofmine / bi-box_regression

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

Bi-box_Regression

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.

Requirements

• Python3
• Pytorch0.4
• torchvision
• tensorboardX

Installation

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

Datasets

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.

Annotations format

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 '.'.

Label encoding file

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

Pretrained Model

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

Training

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>

Reference

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

bi-box_regression's Issues

Can this model be trained on a 1060 graphics card?

How to test the pre-train model?

Is there any python code or instruction to test already exist pre-train model?

Thank you~

This is my result！

这个结果和Repulsion loss一样都变差了，也是奇怪的很！希望作者贴出他的实验结果！

hi~the result of this project can be presented in the command?i have very bad result but i dont konw why

hi~the result of this project can be presented in the command?i have very bad result but i dont konw why

mAP extremely low when testing on Caltech

Hi,

I just found the mAP is very low when testing on Caltech Dataset, it is around 0.03
The code is

optimizer = optim.Adam(retinanet.parameters(), lr=1e-5)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=3, verbose=True)

If I change to SGD or I change lr=1e-4, mAP will always be 0.0
If there anyone has similar problem as mine? What is wrong with my expirement?

BTW, I downsize the caltech to 1/10 of the original size, it takes 35 minutes to run one epoch and I ran 50 epochs. The val data set was chosen from another 1/10 of the original Caltech Dataset. Basiclly the val and test should be very similar, but the mAP on Val is only around 0.3 at 24th epoch.

I have no idea what went wrong.
Thanks!!!

Testing script

I would like to point out that the testing script is some how missing, could you please provide it ?

The use of dataset

I want to use a Caltech Pedestrian dataset for training, but for the transformation of the data set don't correspond with code format and don't know you is how to convert the data set.How to write in CSV, has met the needs of the code format?

Combine Bi-box_Regression and Repulsion loss

Do you try to combine two methods?

pytorch version

Hey, sorry to interrupt, I got a problem with pytorch version and hope you can help me. I am using pytorch 1.0 and for some reason, I can not use pytorch 0.4, so is there any way I can use your code with pytroch 1.0 by changing some parts of your code?

train.txt,val.txt

请问train.txt和val.txt文件格式是怎么样的？