search

bytedance / fc-clip

[NeurIPS 2023] This repo contains the code for our paper Convolutions Die Hard: Open-Vocabulary Segmentation with Single Frozen Convolutional CLIP
Apache License 2.0
285 stars 28 forks source link

About FPS comparison #23

Closed nhw649 closed 8 months ago

nhw649 commented 8 months ago

Hello~ ODISE mentions that "It performs inference for an image (10242) at 1.26 FPS on an NVIDIA V100 GPU and uses 11.9 GB". This does not match the data in table 4. Could you please provide more detailed settings? For example, image size and so on.

cornettoyu commented 8 months ago

Once you run the testing, you will have a testing log file (for both ODISE or FC-CLIP), like the one we provided, which provides the total time for inferring on the whole test set. We report FPS based on number of images / total time to provide a practical estimate on FPS, following Mask2Former

nhw649 commented 8 months ago

ok，thanks.