Inference time

[YJLCV]

Hello, I saw CFNet inference time=0.18 on the kitti benchmark, but I tested the kitti dataset on a GTX1080ti, inference time=0.3, what is your test equipment?

[gallenszl]

Hello, the test equipment is V100 and the inference time of released model CFNet_rvc is 0.22s. For cfnet, you can implement it by adjusting each stage’s (except stage 3) stack hourglass number to be one。 ​

[YJLCV]

@gallenszl thank you for your reply. Can I lose some accuracy or modify some parameters to speed up the model inference time? (I may care more about time performance and can lose some precision)

[gallenszl]

Yes, it's ok

[YJLCV]

@gallenszl Then I would like to ask where the parameters can be modified to speed up the time performance

[YJLCV]

Hello, the test equipment is V100 and the inference time of released model CFNet_rvc is 0.22s. For cfnet, you can implement it by adjusting each stage’s (except stage 3) stack hourglass number to be one。 ​

I hope to be able to modify the parameters during the test phase. If the modification of the parameters is to retrain the model, it will be troublesome.

[gallenszl]

Sorry, you should retrain the model.

[YJLCV]

Hello, the test equipment is V100 and the inference time of released model CFNet_rvc is 0.22s. For cfnet, you can implement it by adjusting each stage’s (except stage 3) stack hourglass number to be one。 ​

Sorry to disturb you, I checked kitti benchmaek and found that you have two algorithms, one is CFNet (time=0.18s, 1 core @ 2.5 Ghz (Python)), the other is CFNet_RVC (time=0.22s, GPU @ 2.5) Ghz (Python)). Would you like to ask what is the difference between the two algorithms? What equipment were used for testing? thank you for your reply!