harlanhong / mm2021-co2-net Goto Github PK

Thanks for your release of THUMOS14. And how to reproduce the results of ActivityNet1.2? Can you provide details?

 数据集不能下载，可以更新一下数据集的下载链接吗？

数据集下载错误提示如下：

可能的原因：
https://answers.microsoft.com/zh-hans/msoffice/forum/all/%E4%B8%8B%E8%BD%BD%E5%88%AB%E4%BA%BA%E9%80%9A/fb0c1cd8-bf17-48e2-819a-c874dad12992

it should be a jpg or png image, shouldn't be a pdf, or it will not shown

In the implementation, I find that function filter_segments is used before evaluation. It seems some predictions are removed by using additional information from "Ambiguous_test".

Hello author, first of all, thank you for your excellent work. Can you tell me how to obtain the JSON result file for proposals? Similar to the template you provided for CO2Net.

CUDA_VISIBLE_DEVICES=0 python test.py --dataset-name ActivityNet1.2 --dataset AntSampleDataset --num-class 100 --path-dataset path/to/ActivityNet1.2 --use-model ANT_CO2 --model-name ANT_CO2_3552 --max-seqlen 60 

test.py not main.py

If I only use the Loss_mil, I get poor result. Can you tell me how did you get the results with only Loss_mil? What is the proposal method? And I see that in you test.py, you wrote a bag_logit, what is it?

As readme suggested, I used your code, data and the same hyper-parameters to train the model on Thumos14 dataset:

However I can only obtain 41.9 mAP@AVG for 0.1:0.7, but the paper reported 44.6, there is about a 2.7% gap, it is a huge gap for this WSL task. Can you explain this?

In paper, about Thumos14 dataset the description is "we use 200 validation videos to train our framework and 213 test videos for testing", 200+213=413, it means that 413 video in total. But when getting the feature of provided from link ,only 412 video features(200 for training ,212 for testing)