inference speed

[arc144]

Hi! First of all thanks for the paper and implementation.

In the paper you mention:

CenterNet, with both center pooling and cascade corner pooling incorporated, reports an AP of 47.0% on the test-dev set, which outperforms all existing one-stage detectors by a large margin. With an average inference time of 270ms using a 52-layer hourglass back-bone [29] and 340ms using a 104-layer hourglass back-bone [29] per image, CenterNet is quite efficient yet closely matches the state-of-the-art performance of the other two-stage detectors.

The inference time you reported is using multi-scale or single-scale setup? I know it is usual to report the single scale time but from your paper it seems you are talking about multi-scale.

Thanks

[Duankaiwen]

@arc144 Hi, the inference time is using single-scale setup