search

wangyu-ustc / LVChat

The official implementation of the paper **LVChat: Facilitating Long Video Comprehension**
12 stars 0 forks source link

The result is inconsistent with Table 4 #3

Open Richar-Du opened 6 months ago

Richar-Du commented 6 months ago

I evaluated LVChat on Egoschema, but the accuracy is below 0.3. I have tried max_num_frm={16, 96, 160}, and all the other hyper-parameters are set as default in inferency.py, but the accuracy is almost the same.

Meanwhile, the accuracy of VideoChat2 is about 0.45, and the max_num_frm={16, 96}.