What is the difference in the video decoding behaviour for decord 0.4.0 and 0.6.0?

[huang-ziyuan]

Hi, thanks for the nice repo for decoding the videos. We have used it since 0.4.0. Recently, we noticed some problems in 0.4.0 for decoding MKV videos, where the decoded frames seem to repeat the first several frames in the video. For example, for a 100-frames video, the 40th to 50th frames decoded might be identical to 10th to 20th frames. It seems to be resolved after 0.5.2.

However, when we have upgraded to 0.6.0, we observe some decoding discrepancies in non-MKV videos. Specifically, for a dataset with avi videos and a model trained on the dataset using decord 0.4.0, the accuracy drops over 1% (from around 45 to 44) when we change the decoder on the test set from decord 0.4.0 to decord 0.6.0.

The experiment is strictly controlled with the only changing factor being the decoder itself. Therefore, I wonder what is the difference in decoding videos for 0.6.0 and 0.4.0, and is the difference in the decoding behaviour a determined pattern that can be reversed in some way? E.g., changing the input index to the 0.6.0 to decode the same set of frames as the 0.4.0.

Thanks again and looking forward to your reply!

[liuyuyuil]

Same accuracy drops found from decord 0.4.0 to decord 0.6.0 in our project.