Detection Speed

[GBJim]

Hi all:

Since this source code is slower than the version on the paper. I want to know exactly the detection speed of this released PVANET.

I tested the detection speed on PVANET and PVANET-compressed on Tesla K80 with images of 1280*720 resolution.

This is what I get:

PVANET --> 5.68 FPS PVANET-compress --> 5.90 FPS

This is about 2.5 FPS faster than Faster-RCNN(VGG16) on my machine.

I want to confirm my detection time are consistent with your detection time. Please give some comments. Thanks! :)

[vj-1988]

I tested the PVANET on TitanX gpu. I am also getting around 5-6 FPS for the images (net.forward() only).

[GBJim]

Thanks @vj-1988 !

[vj-1988]

@GBJim : According to [https://arxiv.org/pdf/1611.08588.pdf] , the PVANET is supposed to run above 20 FPS. I tried batching the images, but found that batching won't work in faster-rcnn frameworks (yet to be implemented). So getting 5-6 fps is not normal.

[GBJim]

@vj-1988: As the author disclaimed on the README.md: This version of py-faster-rcnn is slower than our in-house runtime code (e.g. image pre-processing code written in Python) I am not expecting to get 20 FPS with Python implementation. This is also the reason I need to confirm the detection speed with others :)

By the way, what's the resolution of testing image on paper?

[twmht]

@GBJim

I got 6FPS when testing.

My image resolution is 1920 * 1080

How about yours?

[vj-1988]

My test images were 1280*720 pixels

[twmht]

@vj-1988

The author has claimed that this repo is written in python, thus slower than the paper's.