pretrained action recognition model can not get the score 93.44%,where am I wrong?

I want to test the score of model with the command ‘python3 main.py --dataset ucf101_1 --model se-resnext101-32x4d_vtn_rgbdiff -b128 --lr 1e-5 --seq 16 --st 2 --no-mean-norm --no-std-norm --no-train --no-val --test --pretrain-path …/light_model/se_resnext_101_32x4d_vtn_rgbd_ucf101_s1.pth ’,but I only get the score of 84% ，how can I get the score as the models tell 93.44%``

[vadimadr]

I think the problem is in preprocessing. I guess you used ffmpeg and forgot -q option which controls the quality of generated frames. And by default the quality is poor, resulting in images with visual artifacts.

Use the utils/preporcess_videos.py script it will call ffmpeg with correct options for you.

Btw resnext-vtn-rgb model should give 91.8%.

[minhhoangbui]

@Xiehuaiqi I have the same problem here and I need to clarify your case: You reported that you have a better result with this option

–video-format frames’ which means q=4

However, according to their documents, 4 means the lowest quality. I don’t really get that