What about the test speed on ExtremeNet

[jerrywgz]

Could you please provide the test speed info ?

[AlexeyAB]

@jxingyizhou Hi,

Greate work!

https://www.zpascal.net/cvpr2019/Zhou_Bottom-Up_Object_Detection_by_Grouping_Extreme_and_Center_Points_CVPR_2019_paper.pdf

We use flip augmentation for testing. In our main comparison, we use additional 5⇥ multi-scale (0.5, 0.75, 1, 1.25, 1.5) augmentation. Finally, Soft-NMS [1] filters all augmented detection results. Testing on one image takes 322ms (3.1FPS), with 168ms on network forwarding, 130ms on decoding and rest time on image pre- and post-processing (NMS).
...
Edge aggregation Edge aggregation (Section 4.3) gives a decent AP improvement of 0.7%. It proofs more effective for larger objects, that are more likely to have a long axis aligned edges without a single well defined extreme point. Removing edge aggregation improves the decoding time to
76ms and overall speed to 4.1 FPS.

• Does it mean, that you achieve 4.1 FPS for ExtremeNet (MS) on Tesla V?
• Do you use flip augmentation for SS(single scale) model?
• Does it mean, that you achieve 4.1 FPS for ExtremeNet (MS) and 5x more = 20.5 FPS for ExtremeNet (SS) - use only 1 scale x 2 flip on Tesla V?

ExtremeNet (SS) Hourglass-104 511x511 - 40.2% AP - 55.5% AP50 - 20.5 FPS (V)
ExtremeNet (MS) Hourglass-104 511x511 - 43.7% AP - 60.5% AP50 - 4.1 FPS (V)