dnjs3594 / eigencontours Goto Github PK

Hi,

It seems that the F measure is computed by using batch_f_measure provided by davisinteractive:

But when I run the code, my F measure results are really high. For M=8, the average F score on COCO2017 of my experiments is around 0.84, which is much higher than that of around 0.4 reported in the paper (see Fig. 8 (c)).

In particular, even with M=1, the metric is very high. As a specific example, the following display imgs (left is the approximate mask with cfg.dim=1, right is the ground-truth mask) have a computed batch_f_measure of 1.0. I think there is something wrong since the two masks are so different.

More related questions:

1. I take the SVD of the full contour matrix computed over COCO2017 train, whose size is 360 x 557905, and during the convert step, I apply it to COCO2017 val. Although the paper does not mention how to use train/val, I assume this is the correct way. Can you please confirm this?

2. The paper mentions to compute the F measure, "bipartite matching is performed between the boundary points of a ground-truth contour and its approximated version" and then "F score is defined as the harmonic mean of the precision (P) and the recall (R) of the matching results". In the code this is done by using batch_f_measure provided by davisinteractive between the ground truth masks and approximate masks. Is this equivalent to the way described in the paper?

Looking forward to your reply. Thank you very much!

Very insightful work! I am wondering when will the code for KINS dataset and SBD dataset be released as that for COCO?

Also, is there any plan of releasing code for instance segmentation by embedding eigencontours in YOLOv3?

Thank you so much!

Very much looking forward

Hi, I found your work is very interesting! I have a quick question. It seems that the eigencontours for COCO is computed only over 40000 subsamples based on the following line in the code:

However, in the paper it says that "the proposed eigencontours are determined for all instances in all categories in a training dataset". Since 40k < 110k (size of COCO), which is not "all instances", I wonder which one is correct?

Thanks!

The pkl file is required to perform the first step of preprocessing

preprocessed-data
checkpoint

1

.....

Great work, but I'm still a little confused。 M can control the shape of the object, but how to determine the size and position of the object, so I'd love to learn the rest of the code