search

Dong-JinKim / DenseRelationalCaptioning

Code of Dense Relational Captioning
https://sites.google.com/view/relcap
MIT License
67 stars 13 forks source link

Test on raw images #6

Open Cranooooooo opened 4 years ago

Cranooooooo commented 4 years ago

Hi, very interesting work. How can I test some raw images out of the training dataset?

Dong-JinKim commented 3 years ago

Hello. Thank you for your interest in our work.

I have just added the code "run_model.lua" in our repository.

The instruction can be found in the updated README.md file.

Best regards.

Cranooooooo commented 3 years ago

Hi, there is a question about Raw-Testing. I feed some MSCOCO images to the model, all the output captions tend out to be ''the shadow''. Any clues about how to fix that? Thx~~

4/100 processing image /home/mscoco_img/test2014_jpeg/COCO_test2014_000000018258.jpeg
Results are :
{ captions : { 1 : "the shadow" 2 : "the shadow" 3 : "the shadow" 4 : "the shadow" 5 : "the shadow" 6 : "the shadow" } img : FloatTensor - size: 3x540x720 boxes : CudaTensor - size: 3x4 scores : CudaTensor - size: 3x1 }

Dong-JinKim commented 3 years ago

First, I would like to note that the relational captioning task itself is a challenging task (mAP score). The good news is that we updated our code with our newest version model which shows better performance. I would like to recommend using our newest model (MTTSNet+REM).