search

ChenYingpeng / caffe-yolov3

A real-time object detection framework of Yolov3/v4 based on caffe
472 stars 230 forks source link

yolov3-tiny 检测不出内容 #36

Open Fantastic-Han opened 5 years ago

Fantastic-Han commented 5 years ago

通过转换工具，我只获得了一个tiny-yolo-voc.prototxt 和 tiny-yolo-voc.caffemodel。通过修改detectnet.cpp（减少了一个blob），yolo_layer.cpp（改变循环参数），能够进行检测，但是没有检测到任何东西，请问这是怎么回事呢？

passion3394 commented 5 years ago

遇到同样的问题，作者能出来解释一下嘛，多谢！

ChenYingpeng commented 5 years ago

Please check note 4 Yolov3-tiny: Caffe can not duplicate the layer that maxpool layer (params:kernel_size = 2,stride = 1),so rewrite max_pool_1d function for recurrenting it. I convert yolov3-tiny.prototxt to yolov3-tiny-1.prototxt and yolov3-tiny-2.prototxt.Detail check Example 4:yolov3-tiny.

passion3394 commented 5 years ago

yes, I followed the note 4 to split the original yolov3-tiny prototxt to two prototxt, the result is very bad.

passion3394 commented 5 years ago

the original Darknet model works normally, but the converted yolov3-tiny-1.prototxt and yolov3-tiny-2.prototxt can only detect a point, which is the center of the object.

Fantastic-Han commented 5 years ago

遇到同样的问题，作者能出来解释一下嘛，多谢！

yes, I followed the note 4 to split the original yolov3-tiny prototxt to two prototxt, the result is very bad.

How do you split the original yolov3-tiny prototxt to two prototxt? I don't understand the specific method very well.

Fantastic-Han commented 5 years ago

Please check note 4 Yolov3-tiny: Caffe can not duplicate the layer that maxpool layer (params:kernel_size = 2,stride = 1),so rewrite max_pool_1d function for recurrenting it. I convert yolov3-tiny.prototxt to yolov3-tiny-1.prototxt and yolov3-tiny-2.prototxt.Detail check Example 4:yolov3-tiny.

I am not very clear how to split the original yolov3-tiny prototxt to two prototxt？Can you explain it briefly?

passion3394 commented 5 years ago

@dd110hh110 please refer the link on the readme of this repo, the author has published the two spited prototxt in the Baidu net disk.

https://pan.baidu.com/s/1yiCrnmsOm0hbweJBiiUScQ

Fantastic-Han commented 5 years ago

@dd110hh110 please refer the link on the readme of this repo, the author has published the two spited prototxt in the Baidu net disk.

https://pan.baidu.com/s/1yiCrnmsOm0hbweJBiiUScQ

在您的提示下，我参考了里面的prototxt文件，终于可以检测出视频内容了，非常感谢您的意见。 但是检测效果不太理想，漏检率还是很高的，但是检测速度在tx2上有20帧左右，速度还可以。 对比了下直接使用darknet，差距还是很大的。

passion3394 commented 5 years ago

@dd110hh110 我觉得作者的这种实现方式可能有点问题，期待作者能改进