Orientation Estimation

[trungpham2606]

@yangli18 What a great work. I just saw your work few days ago, read your paper but when looking into your codes, I cant see where Projection loss and Wrist loss are ?

[yangli18]

@trungpham2606
As both of them are defined as the smooth L1 loss, they are calculated along with the location loss L_loc in multibox_loss.py.

[trungpham2606]

@yangli18 oh, thank you. In addition, can you give me information about the loc_loss and cls_loss after finishing your training session. It seems that your log file doesn't contain the final results.

[yangli18]

@trungpham2606
I just updated the training log. It is corresponding to the trained model in weights dir.

[trungpham2606]

@yangli18 oh I see it. Thank you for your prompt response. I had tried your network with my own dataset and the results are immensely good !

[yangli18]

@trungpham2606 I'm glad it helped! 😄

[trungpham2606]

@yangli18 Can you tell me how much time the model needs for showing results on 1 image ( prediction time + display time). In my PC, the amount of time for 1 image is approximately 1sec ( I was using pure Python nms).
Thank in advance.

[yangli18]

@trungpham2606
We tested the average detection time (including NMS) on the Oxford hand dataset, and it's about 7 ms/image using Titan X (Maxwell). We didn't count the display time and you should test it yourself on your platform. The NMS written in c/cuda is also provided and it's much faster.

[trungpham2606]

@yangli18 Yeah, I also guessed that it took most time in display part while I was using pure Python nms ( i cant install nms c/cuda built on windows 10, I will try cython version).
Thanks so much for helping me ^^.

[yangli18]

@trungpham2606 My pleasure! Thanks for your interest.