Hi, thanks for releasing the code.
I have a question regarding the pre-computation stage of question-frame relevance scores.
From my understanding, the implementation samples video frames at 1 FPS and computes the similarity between each sampled frame and the question. For long videos or large-scale datasets, this preprocessing step can be quite time-consuming.
I was wondering:
- Did you impose any upper limit on the number of sampled frames per video during pre-computation?
- For very long videos, did you keep all 1-FPS sampled frames, or did you apply any additional filtering/downsampling strategy?
- Approximately how long did the relevance pre-computation take for the datasets used in the paper?
Any details about the practical setup would be greatly appreciated.In my own implementation, a 10-minute video sampled at 1 FPS already produces around 600 frames, and computing CLIP/BLIP similarities for every frame becomes a noticeable bottleneck. Therefore, I'm particularly interested in whether there was a frame-budget constraint (e.g., maximum number of frames per video) in your experiments.
Thanks!
Hi, thanks for releasing the code.
I have a question regarding the pre-computation stage of question-frame relevance scores.
From my understanding, the implementation samples video frames at 1 FPS and computes the similarity between each sampled frame and the question. For long videos or large-scale datasets, this preprocessing step can be quite time-consuming.
I was wondering:
Any details about the practical setup would be greatly appreciated.In my own implementation, a 10-minute video sampled at 1 FPS already produces around 600 frames, and computing CLIP/BLIP similarities for every frame becomes a noticeable bottleneck. Therefore, I'm particularly interested in whether there was a frame-budget constraint (e.g., maximum number of frames per video) in your experiments.
Thanks!