daybreak-u / yolo-face-with-landmark

yoloface大礼包 使用pytroch实现的基于yolov3的轻量级人脸检测（包含关键点）

yolo-face-with-landmark

实现的功能

• 添加关键点检测分支，使用wing loss

Installation

Clone and install

1. git clone https://github.com/ouyanghuiyu/yolo-face-with-landmark
2. 使用src/retinaface2yololandmark.py脚本将retinaface的标记文件转为yolo的格式使用,
3. 使用src/create_train.py 创建训练样本

训练

python train.py --net mbv3_large_75 --backbone_weights \
./pretrained/mobilenetv3-large-0.75-9632d2a8.pth --batch-size 16 

测试

python evaluation_on_widerface.py
cd widerface_evaluate
python evaluation.py

demo

python demo.py

精度

Widerface测试

• 在wider face val精度（单尺度输入分辨率：320*240）

方法	Easy	Medium	Hard	Flops
Retinaface-Mobilenet-0.25(Mxnet)	0.745	0.553	0.232
mbv3large_1.0_yolov3(our)	0.861	0.781	0.387	405M
mbv3large_1.0_yolov3_light(our)	0.856	0.770	0.370	311M
mbv3large_0.75_yolov3(our)	0.853	0.778	0.382	334M
mbv3large_0.75_yolov3_light(our)	0.845	0.766	0.365	240M
mbv3samll_1.0_yolov3(our)	0.798	0.696	0.3	185M
mbv3small_1.0_yolov3_light(our)	0.759	0.662	0.300	91M
mbv3samll_0.75_yolov3(our)	0.768	0.673	0.305	174M
mbv3small_0.75_yolov3_light(our)	0.754	0.647	0.291	80M

• 在wider face val精度（单尺度输入分辨率：640*480）

方法	Easy	Medium	Hard
Retinaface-Mobilenet-0.25(mxnet)	0.879	0.807	0.481
mbv3large_1.0_yolov3(our)	0.900	0.882	0.707
mbv3large_1.0_yolov3_light(our)	0.900	0.874	0.683
mbv3large_0.75_yolov3(our)	0.886	0.871	0.694
mbv3large_0.75_yolov3_light(our)	0.881	0.862	0.678
mbv3samll_1.0_yolov3(our)	0.856	0.827	0.602
mbv3small_1.0_yolov3_light(our)	0.847	0.807	0.578
mbv3samll_0.75_yolov3(our)	0.841	0.815	0.584
mbv3small_0.75_yolov3_light(our)	0.832	0.796	0.553

ps: 测试的时候,长边为320 或者 640 ,图像等比例缩放

测试

References

• yolov3

yolo-face-with-landmark's Issues

这套代码有keras版本的吗

Train a model with 64 or 98 key points

Thank you very much for your innovative work, if I want to train a model with 64 or 98 key points. What other modifications are needed besides modifying point_num. Please provide help, thanks!

How to train my own dataset?

Hi!It is my honor to see your codes,I want to know how to train my own fish datasets instead of people datasets.Which codes should I change?I would appreciate it if you could tell me!Thanks!

检测框的置信度

标签文件示例

你好，可不可以给一下标签文件 的示例呢？非常感谢

torch模型转caffe

十分感谢大佬的代码，请问下，这个有尝试过pytorch的模型转成caffemodel么，好不好操作，损失大么？

ncnn inference why low accuracy?

Hi,

I am trying to evaluate all models in ncnn. But even with ncnn converted model mbv3_small_1.bin/param the detection accuracy is much lower. I would like to know did you verified from your end that ncnn inference also gives the same result as this pytorch inference. If required I can provide detection results with eval numbers. Just for reference I have attached below ncnn detection result as well as python pytorch detection result in a sample image with same input size.

pytorch inference:

• detected faces: 864
  

ncnn infernence:

• detected faces: 534
  

请问训练人脸检测默认设置'--single-cls'为false吗？

convert onnx

mbv3_small_75_light

请问大佬如何将pytorch模型转到ncnn的，有没有脚本可否分享一下？谢谢

复现train.py的一个报错。

`
Saving labels to /home/ps/superb/yolo-face-with-landmark/datas/yololandmark_wider_train.npy for faster future loading
Image sizes 320 - 640 train, 640 test
Using 8 dataloader workers
Starting training for 250 epochs...

 Epoch   gpu_mem      GIoU       obj       cls      land     total   targets  img_size        lr

0%| | 0/805 [00:00<?, ?it/s]Traceback (most recent call last):
File "train.py", line 365, in
train(hyp) # train normally
File "train.py", line 265, in train
loss, loss_items = compute_loss(pred, targets, model,point_number=hyp['point_num'])
File "/home/ps/zhouyizhi/yolo-face-with-landmark/utils/utils.py", line 238, in compute_loss
tcls, tbox, tpoint ,indices, anchors = build_targets(p, targets, model,point_number=point_number) # targets
File "/home/ps/zhouyizhi/yolo-face-with-landmark/utils/utils.py", line 342, in build_targets
no_train_index = torch.where(targets == -1 )
TypeError: where() missing 2 required positional argument: "input", "other"
`
请问下作者，这个错误该如何解决。

请问关键点数据库从哪里下载？

谢谢

How to prepare data for evaluation?

Is it the same as train data?

关键点不可见

pytorch 和torchvision 版本

你好，pytorch 和torchvision 版本是多少呢，我训练报下面的错
TypeError: _resolve_type_from_object(): incompatible function arguments. The following argument types are supported:
1. (arg0: object, arg1: torch._C._jit_tree_views.SourceRange, arg2: Callable[[str], function]) -> torch._C.Type

Invoked with: typing.Union[int, NoneType], None, <function try_ann_to_type..fake_rcb at 0x7f52e2a0de18>
谷歌搜了一下说是版本问题

博主，请问有具体来说明解释哪里做了改动吗？

博主，请问有具体来说明解释哪里做了改动吗？比如对那些文件进行了改写之类的

如何用opencv或者mnn推理呢？