search

OpenGVLab / video-mamba-suite

The suite of modeling video with Mamba
MIT License
216 stars 21 forks source link

Discrepancy in Video Temporal Grounding Task Results on QVHighlights Dataset #13

Open Summer-seu opened 3 months ago

Summer-seu commented 3 months ago

Thank you for your excellent work.

For the Video Temporal Grounding task, I trained the model using your provided "bash scripts/qvhl_pretrain_mamba.sh" command. However, the metrics I obtained were about 12% lower than your reported results on the QVHighlights dataset.

Does your approach require pretraining similar to UniVTG? I couldn't find any (pretrained) checkpoints for the Video Temporal Grounding task in your model zoo.

cg1177 commented 3 months ago

In our experiments, we did not perform any pretraining process.

Summer-seu commented 3 months ago

Could you pls provide the checkpoints of Video Temporal Grounding task? Thanks.

cg1177 commented 3 months ago

Could you pls provide the checkpoints of Video Temporal Grounding task? Thanks.

Sure, we can provide the checkpoints for the Video Temporal Grounding task. However, we will need some time to re-train the models.

Summer-seu commented 3 months ago

Thanks so much!