search

iwatake2222 / play_with_tensorrt

Sample projects for TensorRT in C++
Apache License 2.0
190 stars 34 forks source link

pj_tensorrt_seg_paddleseg_cityscapessota question #4

Closed LutaoChu closed 2 years ago

LutaoChu commented 2 years ago

Project Name

pj_tensorrt_seg_paddleseg_cityscapessota

Issue Details

Awesome work！ What does the picture in the lower-left corner of your demo mean? How is it generated?

image

iwatake2222 commented 2 years ago

Thanks!

I colored each pixel weighted by scores of all 19-classes, while the original PaddleSeg code outputs only one class with the highest score for each pixel using ArgMax. For instance, scores of the pixel near border b/w person and bicycle can be like {Person = 0.49, Bicycle = 0.51}, and Bicycle is selected for the pixel. It might be correct or wrong. So, instead of this normal way, I colored 0.49 of person's color and 0.51 of bicycles's color. This image may have more meaning and be useful for some purposes (I don't have any idea so far, though...)

Besides, it LOOKS nice, isn't it? ;)

Note: although I wrote "score" above, the model seems output logit value so the range is not [0, 1.0]. I did some calculation to convert to weight.

LutaoChu commented 2 years ago

Thanks for your explanation. I haven't thought of any use, but it does sound interesting. This seems to have more details than using argmax. Is it superimposed with the original image?

iwatake2222 commented 2 years ago

No, I colored based on a black image (cv::Mat::zeros ). This is an interesting part. I used only the output data of the model, still I can easily imagine the original image. Also, it looks this image contains necessary and sufficient information to drive !

LutaoChu commented 2 years ago

Yes, very interesting discovery. Waiting for us to explore its usage scenarios!

iwatake2222 commented 2 years ago

This issue was closed because it has been inactive.