d3s on davis2016

[riversci]

Hi, I think d3s is an excellent work which combined both tracking and segmentation. I like it very much. However, when I tested the pretrained model on davis2016, I got results as follow.

it's not as same as the paper.

I am confused. Can you give me some advice?

[alanlukezic]

One possible reason for different scores is different hardware. We observed that the performance can differ due to different GPUs or even library versions.
The other possible reason is tracker initialization. It should be initialized as:
tracker.initialize(img, gt_rect, init_mask=gt_mask)
where gt_rect is an 8-element vector of the rectangle corners: [x0, y0, x1, y1, x2, y2, x3, y2].

[riversci]

Thank you for your answer. But How you get the gt_rect from gt_mask? Can you share some details?

[laisimiao]

@alanlukezic could you write a guide doc to help us evaluate D3S on davis such vos datasets?

[laisimiao]

@riversci how you test on davis2016, could you give some guidences?