Datasets:
Missing Video_Ret files
I ran data download from the official github repo, and it downloaded the MMEB-v2 benchmark for me. However, when I try to evaluate the model, I get the following error. I check the directory structure of the benchmark here and I see that the video retrieval files are missing. Can you please help me with this?
[rank2]: FileNotFoundError: File vlm2vec_eval/video-tasks/videos/video_ret/MSR-VTT/video7020.mp4 does not exist
+1
@Charliieeee
@VarunGumma
Thanks for your question. I want to make sure I fully understand your situation. The existing video evaluation uses the pre-sampled frame files. Could you let me know why you are using the .mp4 files here?
The current evaluation logic is as follows: if the frame files are available, the evaluation uses them directly. If they are not available (e.g., for users who prefer to perform their own frame sampling), it will fall back to the raw video files.
Just FYI, not sure whether your question is the same as this one. You could also refer to my answer there. Let me know if there is any question.
https://github.com/TIGER-AI-Lab/VLM2Vec/issues/159
@ziyjiang
+1 on my side. Seems video_ret.tar.gz has some hard-coded unzip destination folder. For MSR-VTT dataset I'm getting all the frames placed in "$BASE_DIR/video-tasks/frames/video_ret/data/ziyan/video_retrieval/MSR-VTT/frames" after decompression. That explains why evaluation logics cannot find proper frames and turn to look for mp4/avi files instead.
For me moving everything inside that folder to "$BASE_DIR/video-tasks/frames/video_ret/MSR-VTT" as configured by video.yml works, same for the other video_ret datasets.
Could you double check that file please?