Why I run peleeNet to detect video and cost 78.5ms per image while SSD cost only 40.5ms per image about pelee

[ujsyehao]

You can use ./build/tools/caffe time --model */deploy.prototxt --gpu 0 command to get time cost on each layer.
Running time can be affected by batch size, pre-process(image preprocess), post-process(NMS), code implementation, So we use MACs to evaluate computational cost which can intuitively reflect the running time.

[zhangjinsong3]

@ujsyehao I run it in tensorRT environment, and it run quite faster than SSD,thanks for your contribution again.

[ujsyehao]

Can you post your test speed?

[zhangjinsong3]

@ujsyehao in tesla P4，with tensorRT soupport ,peleeNet can run in 30ms with the batch_size = 8. While SSD+VGG16 (we have made some compression ) needs 90ms.

[ujsyehao]

@zhangjinsong3 Thank you, tesla P4 is P40?

[zhangjinsong3]

@ujsyehao Just p4,not p40.

[ujsyehao]

@zhangjinsong3 okay, according to graphics card's tflops, the speed is fast.