the performace without deformable conv?

[meteorshowers]

can you provide the performace without deformable conv?

[meteorshowers]

Moreover, I test your model with normal convolution,100ms on titanxp is the runtime right?

[haofeixu]

You can refer to the Table 1 in our paper for the performance without deformable conv.

May I know your test settings (e.g., resolution)? Our inference time is measured by inference.py with --count_time activated.

[meteorshowers]

Thanks 360*1280 I do not use the deformable conv.

[haofeixu]

Are you using our code inference.py?

[meteorshowers]

no

[meteorshowers]

why do not eval on the 1/8 resolution with stereonet config? It would be faseter

[haofeixu]

There is a trade-off between accuracy and speed. If speed is a priority, you can use the StereoNet-AA model.

[meteorshowers]

If anyone test the inference time on titanxp, tell me, thanks

[haofeixu]

We have uploaded our timing scripts, you should use them to measure the inference time.

[meteorshowers]

118.24091136 gflops 3.175282 M double check parameter 3.175282 M time 114.790651798 ms input 360*1260 device titian xp I think v100 may 1.5 times faster than titianxp, it is normal

[wangzhao88]

118.24091136 gflops 3.175282 M double check parameter 3.175282 M time 114.790651798 ms input 360*1260 device titian xp I think v100 may 1.5 times faster than titianxp, it is normal

您好，能否支援一份不带dconv的代码？

[yanzhou-li]

can you share me the code without @@deformable

can you provide the performace without deformable conv?

can you share me the code without deformable,thank you! @meteorshowers