The DAVIS Challenge on Video Object Segmentation @ CVPR 2018 Forum

Go back to competition Back to thread list Post in this thread

> How should the segmention outputs be?

hello, i am new to multi instance segmentation,i find that the annotations is something like the pixel would be 128 in some of the png RGB channel, howerver, there is no limit in the number of categories in total, and in different image the color order may change for same category, so how should i label my segmentation model,originally i perform softmax on the bottleneck which contains N+1 channel,but in this task i was stuck at this.

Posted by: naiven @ May 17, 2017, 11:21 a.m.

Hi,

have a look at https://github.com/jponttuset/davis-matlab and also https://github.com/jponttuset/davis-matlab/blob/master/helpers/create_submission_zip.m
This will give you some hints about the format.
And if you want to understand the color coding have a look here:
https://github.com/jponttuset/davis-matlab/blob/4b71631dbd5406629e61c75c2117eed78afab3b5/db_util/auxiliar/pascal_colormap.m
If you read the image in the right way (e.g. with PIL.Image), then you will directly get the object ids instead of the colors, since the image is created with an indexed palette.
I hope, that helps you.

Best,

Paul

Posted by: voigtlaender @ May 24, 2017, 7:42 a.m.

also note, that DAVIS is not about categories, but about instances. So the same color/id can be used for different types of objects in each sequence! And you're right, the number of annotated objects is different between sequences

Posted by: voigtlaender @ May 24, 2017, 7:44 a.m.
Post in this thread