Require LVIS v0.5 for research

Inyoung Yoon

unread,

Jan 9, 2023, 12:59:22 AM1/9/23

to lvis-d...@googlegroups.com

Hi, I'm In Young Yoon, a researcher of SPALAB at Hanyang University.

I'm researching long-tailed recognition.

I need LVIS v0.5, because it has experimented with previous methods, but the website doesn't provide it.

So can you provide LVIS v0.5 if possible?

Ross Girshick

unread,

Jan 9, 2023, 6:54:02 PM1/9/23

to LVIS Dataset

Hi Young Yoon,

We discourage the use of LVIS v0.5 -- you should use v1.0 instead. In order to help normalize the research community on v1.0 we have intentionally made v0.5 hard to obtain. If the previous method doesn't have results on v1.0, we suggest that you run that method on v1.0 of the dataset. It may also suggest that it's an outdated baseline and other comparisons are now more relevant.

Best,

Ross

Wilson Lee

unread,

Nov 14, 2023, 3:33:51 AM11/14/23

to LVIS Dataset

Hi Dr. Girshick,

I am Wilson Lee, a researcher from Tsinghua University.

I notice that in your paper "segment anything model", you provide results on lvis val v0.5 when evaluate the results of points prompt (see Table 7). Could you please provide the results of v1.0 or provide me the LVIS v0.5 version to reproduce this results if possible?