mengyuanchen21 / eccv2022-delu Goto Github PK

Hi，authors.Thanks for your wonderful work！
I carried out the experiment on thmos14 according to the code you provided. The machine is nvidia GeForce 3090. I tried random seeds 0-4, but the highest one is still lower than your accuracy in the paper. Can you give me some suggestions.The average precision of random seeds from 0 to 4 is 45.7, 45.6, 45.4, 45.6, 46,1 respectively.

Hi Authors,
Thanks for your wonderful work！
I create a clean environment and follow the provided code to train the model, but the results on THUMOS14 are much lower than those reported in the paper.
Iteration: 4950, Loss: 2.23683
Testing test data point 0 of 212
Testing test data point 42 of 212
Testing test data point 84 of 212
Testing test data point 126 of 212
Testing test data point 168 of 212
Testing test data point 210 of 212
Number of ground truth instances: 3358
Number of predictions: 36474
Classification map 93.916390
map @ 0.1 = 67.505 ||map @ 0.2 = 62.470 ||map @ 0.3 = 53.532 ||map @ 0.4 = 45.423 ||map @ 0.5 = 37.230 ||map @ 0.6 = 26.003 ||map @ 0.7 = 13.898 ||map @ 0.8 = 7.284 ||map @ 0.9 = 2.063
mAP Avg ALL: 35.045
MAX map @ 0.1 = 69.091 ||MAX map @ 0.2 = 63.765 ||MAX map @ 0.3 = 54.554 ||MAX map @ 0.4 = 45.960 ||MAX map @ 0.5 = 38.204 ||MAX map @ 0.6 = 26.796 ||MAX map @ 0.7 = 14.790 ||MAX map @ 0.8 = 7.460 ||MAX map @ 0.9 = 2.137
mAP Avg 0.1-0.5: 54.3147951833503, mAP Avg 0.1-0.7: 44.73712870057078, mAP Avg ALL: 35.8617997448392
------------------pid: 1102141--------------------
I have the same configuration for the environment as the suggested. Can you give me some advice?

Hello author, first of all, thank you for your excellent work!
I wanted to replicate the code using Activitynet 1.2, but found that there was no corresponding .sh code file. So I copied the THMOS14 running file and wrote an Activitynet 1.2 file, but encountered many problems. Is it correct to run the Activitynet dataset like this

I really want to know the algorithmic details of the article, but the data link given cannot be opened。 Could you give me another link such as baidu link or others?

Hello, author. I would like to ask if the code for DELU can only be run through a single gpu？