In this project, you can enjoy yolo-v2, yolo-v3, tiny-yolo-v2 and tiny-yolo-v3. What I have to say is that I don't try to 100% reproduce official YOLO, because it is really difficult and I have not much computation resource. Therefore, I did some simplification. For example:

For objectness, official YOLO firstly computes iou between all predicting bboxes and groundtruth bboxes, then pick up the one whose iou is the largest and set the learning target of objectness as the iou value. But what I do is to set objectness target as 1 if there is a object, otherwise 0. My method is really easy, right? I'm so lazy. HAHA!

In a word, this project is an exercise, and I prefer to call it a lazy version of YOLO. HAHA!! If you guys are interested in the authentic YOLO reproduced by PyTorch, this project might not suit you.

Good news again !!!

In this update, I add tiny-yolo-v2 and tiny-yolo-v3 that are both designed by myself. My tiny-yolo-v2 gets 61.6 mAP and 117 FPS on a GTX-1660ti GPU and tiny-yolo-v3 gets 71.3 mAP and 62.5 FPS on a GTX-1660ti GPU. There is nothing new but just tiny. So, have fun !

Good news !!!

In this update, I add yolo-v3 model which gets 81.0 mAP ( with 416 input, and no multi-scale training trick ) on VOC2007-test. For more details, you could see code files including models/yolo_v3.py and tools.py.

In addition, you can replace darknet-53 with darknet-19 as the backbone of yolo-v3, and this model can get 78.6 mAP on VOC2007-test.

I really enjoy yolo. It is so amazing! So I try to reproduce it. And I almost got it!

But !! I don't plan to one hundred percent reproduce it, so my own yolo-v2 is a little different with origin version. Because all the tricks used in yolo-v2 may not ne suitable for other tasks.

It is known to us that reproducing the model is easy while it is hard to reproduce the results shown in the paper.

As the Chinese new years(2020-1-24) is comming, I have no more energy to adjust my yolo-v2 which means my model can't get the same result as origin yolo-v2 (76.8 mAP with 416 and 78.6 with 544). If you really care about this point, my project can't satisfy you. And there are many other excellent projects reproducing origin yolo-v2, so just consider to clone them and try on your task.

Before I tell you guys how to use this project, I must say something about difference between origin yolo-v2 and mine:

• For objectness, I just regard it as a 0-1 bernoulli distribution (Origin yolo-v2 set IoU as the objectness label.). If there is an object, the label is 1, else it is 0. And the model uses sigmoid function and MSE loss function.

• For class prediction, I use cross-entropy funtion while origin yolo-v2 used MSE to regression it. I really can't understand why it used MSE for class. If anyone knows that, please tell me. Thanks a lot!

• For data augmentation, I copy the augmentation codes from the https://github.com/amdegroot/ssd.pytorch which is a superb project reproducing the SSD. If anyone is interested in SSD, just clone it to learn !(Don't forget to star it !)

My yolo-v2 got 75.2 mAP with 416 input size, lower than origin yolo-v2 that got 76.8% mAP with the same image size. With 608 input size, my yolo-v2 gets 77.2 mAP. This is maybe because that there are two tricks that didn't work:

1. hi-res classifier

The darknet19-448 has been trained and got 75.52 top-1 acc. But it doesn't bring any improvement for my model.

2. multi-scale train

It doesn't work, too. Although it does make my model robust to more input size (from 320 to 608), it didn't bring any increase on mAP with 416 resolution.

My loss function and groundtruth creator both in the tools.py, and you can try to change some parameters to improve the model.

• Pytorch-gpu 1.1.0
• Tensorboard 1.14.
• python-opencv, python3, Anaconda3-5.1.0

As for now, I only train and test on PASCAL VOC2007 and 2012.

I copy the download files from the following excellent project: https://github.com/amdegroot/ssd.pytorch

I have uploaded the VOC2007 and VOC2012 to BaiDuYunDisk, so for researchers in China, you can download them from BaiDuYunDisk:

Link：https://pan.baidu.com/s/1tYPGCYGyC0wjpC97H-zzMQ

Password：4la9

You will get a VOCdevkit.zip, then what you need to do is just to unzip it and put it into data/. After that, the whole path to VOC dataset is data/VOCdevkit/VOC2007 and data/VOCdevkit/VOC2012.

# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2007.sh # <directory>

# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2012.sh # <directory>

I copy the download files from the following excellent project: https://github.com/DeNA/PyTorch_YOLOv3

Just run data/scripts/COCO2017.sh

To run:

python train_voc.py

You can run python train_voc.py -h to check all optional argument

To run:

python test_voc.py --trained_model [ Please input the path to model dir. ]

To run:

python eval_voc.py --train_model [ Please input the path to model dir. ]

Finally, I have tried to train my yolo-v2 on MSCOCO-2017 datatset, but I didn't get a good result. My yolo-v2 got 31.5 AP50 on MSCOCO-valid dataset and its visualization results are a little poor. I haven't address this problem , but what I know is that MSCOCO is really a challenging dataset !