search

WeitaiKang / SegVG

[ECCV 2024] SegVG: Transferring Object Bounding Box to Segmentation for Visual Grounding
41 stars 2 forks source link

How large is the computation? #2

Closed KevinfromTJ closed 1 month ago

KevinfromTJ commented 4 months ago

I wonder if you could share more details about the training cost (GPUs and days)

WeitaiKang commented 4 months ago

Sure,

For costs, I just compared GFLOPs and training time with VLTVG. I used 4*A6000 GPUs with the same environment to test the training time required to achieve the best performance on ReferItGame. Btw, I remember VLTVG has less cost than TransVG and QRNet. Results are as below. It is affordable I think.

Model GFLOPS (G) Training Time
VLTVG 69.87 16 hours
SegVG 73.48 28 hours