Closed abel71 closed 2 years ago
Hi, thank you for your interest in our paper! So first thing you should know is that the results reported in the paper are using the official Matlab toolkit provided by the authors. The coco metric is only a hint as to if we're improving are results or not. The main reason they differ (in my opinion) is that the "ignore regions" are automatically handled in the Matlab script. Also, the mAP calculation might be slightly different, I am not sure. Good luck!
Hi, I am currently using the official code CenterNet to train the UA-DTETRAC dataset and have seen that your training accuracy using this method is 83%.But my training accuracy is about 10% worse than yours.I would like to ask you about the config file settings during the training process, such as the learning rate and the selection of the backbone network or other changes.In addition,would you like send me the log txt.Many thanks for your support.
Hi, are you using the official UA-DETRAC Matlab evaluation tool? Because the results are better with it than the COCO metric. The parameters for my experiments can be found at: https://github.com/hu64/FFAVOD/tree/master/ffavod-experiments. Other values are set to default, which can be found at: https://github.com/hu64/FFAVOD/blob/master/src/lib/opts.py
Hi @Yubzsz are you able to sign up and login to UA-DETRAC website? I always getting Whoops, looks like something went wrong.
while signing up on the website. And if you don't mind, could you share the UA-DETRAC annotations and its evaluation tools for me? Thanks in advance
Hi @Malikanhar I download from https://pan.baidu.com/s/130w_33XEnwTjP3OOqdkfHA,Extraction code:4rgd. I didn't have the evaluation tools either.
Thank you very much! @hu64 I use the hourglass to train,and achieve the same effect.
Hi, I have a hard time reproducing the results reported in your repo/paper using the provided pre-trained models. On the UAVDT test dataset (using the same split you shared /val-1-on-30.json), I attempted to evaluate the shared per-trained model (spotnet2_vid_uavdt.pth), but I was not able to obtain the same mAP as reported in the paper.
Is there anything else I could do to reproduce the results.