zerollzeng / tensorrt-zoo

openpose, yolov3 with tiny-tensorrt
86 stars 25 forks source link

about the confindence of object #14

Closed guods closed 4 years ago

guods commented 4 years ago

I test YOLOV3 model with TensorRT7.0 (float16), I found that the confindence of all detetected objects is greater than 0.5(no confidence filtering); do you know the reason?

zerollzeng commented 4 years ago

https://github.com/zerollzeng/tensorrt-zoo/blob/dd6381f82cf35cac2c2cb4feb935a0b447e0c128/yolov3/YoloV3.cpp#L141

and

https://github.com/zerollzeng/tensorrt-zoo/blob/dd6381f82cf35cac2c2cb4feb935a0b447e0c128/yolov3/YoloV3.cpp#L88-L94

maybe take a look at this? I forgot the details of nms :)

guods commented 4 years ago

I know do nms; but before nms , the confindence of all detetected objects has been greater than 0.5.

zerollzeng commented 4 years ago

https://github.com/zerollzeng/tiny-tensorrt/blob/69e64d10270b518c1356b50d9dff1553e3d8c1ce/plugin/YoloLayerPlugin/YoloLayerPlugin.hpp#L22

and

https://github.com/zerollzeng/tiny-tensorrt/blob/69e64d10270b518c1356b50d9dff1553e3d8c1ce/plugin/YoloLayerPlugin/YoloLayerPlugin.cu#L283-L300

Just change value of IGNORE_THRESH if you need, note that their are gpu and cpu forward of yolo layer plugin

guods commented 4 years ago

Thanks for your reply, I has been solved it.