search

clovaai / EXTD_Pytorch

Official EXTD Pytorch code
MIT License
187 stars 60 forks source link

The result of the evaluation time doesn't be reproduced. #15

Closed vujadeyoon closed 4 years ago

vujadeyoon commented 4 years ago

Hello, authors. Thanks for releasing the official code for EXTD.

I've tried to reproduce the result of evaluation time in Fig. 6 using the official codes (e.g. commented codes of demo.py and blackpink.jpg). However, I failed to get a similar result on my two workstations. The first one is NVIDIA DGX-1 (4 V100) with PyTorch 1.2, and the last one consists of i7 and 4Titan XP with PyTorch1.2. Of course, I use a single GPU, not multiple GPUs to testing an image.

I wonder how you measure the elapsed time like Fig. 6. I think the number of faces in an image affects the total inference time because of the NMS as far as you know. Thus, I am not sure you really use the blackpink.jpg to get result of the Fig. 6. Also, I am not sure the configuration parameters in ./data/config.py are utilized for it. (The speed can be slightly increased by changing the parameters, but it has not yet reached the reported performance.)

Can I just get similar results like Fig.6 using only the given codes and an image?;)

I am looking forward to hearing from you.

Thanks.

vujadeyoon commented 4 years ago

Now, I know the backbone network is different between codes and papers. I close this issue.